<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Matias Martinez</title>
    <description>The latest articles on DEV Community by Matias Martinez (@matias_martinez_185d9e0ee).</description>
    <link>https://dev.to/matias_martinez_185d9e0ee</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F2610948%2F367b842d-2a5b-47fc-abb1-526da837ee4a.jpg</url>
      <title>DEV Community: Matias Martinez</title>
      <link>https://dev.to/matias_martinez_185d9e0ee</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/matias_martinez_185d9e0ee"/>
    <language>en</language>
    <item>
      <title>Web Answer on Bedrock AgentCore: Cited Web Answers for Your Agents in One Tool Call</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Tue, 10 Nov 2026 14:45:56 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/web-answer-on-bedrock-agentcore-cited-web-answers-for-your-agents-in-one-tool-call-36nd</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/web-answer-on-bedrock-agentcore-cited-web-answers-for-your-agents-in-one-tool-call-36nd</guid>
      <description>&lt;p&gt;Giving an agent access to the web sounds simple until you ship it: pick a third-party search API, feed results to the model, let it decide which pages to fetch, scrape and trim them to fit the context window, synthesize, and (if you remembered) cite sources. Every step adds tokens, latency, another API key to rotate, and another vendor in your risk review.&lt;/p&gt;

&lt;p&gt;On November 5, 2026, AWS announced &lt;strong&gt;Web Answer on Amazon Bedrock AgentCore&lt;/strong&gt;, a managed AgentCore Gateway connector that takes a natural-language question and returns a written answer where every claim is tied to a source URL, in a single MCP tool call. It builds on Web Search, which landed in June and returns ranked results for your own model to process. This post covers how Web Answer differs from Web Search, how to wire it up with the CLI and Strands Agents (including from EKS), the filtering model, pricing, and what to check before you put it in front of users.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Managed Gateway connector (&lt;code&gt;connectorId: "web-answer"&lt;/code&gt;) that returns a &lt;strong&gt;synthesized, cited answer in one tool call&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Under the hood, an AWS-managed agent plans searches, runs several in parallel, reads key pages to verify facts, and searches again when evidence is thin.&lt;/li&gt;
&lt;li&gt;Searches an &lt;strong&gt;Amazon-operated web index plus Amazon's Knowledge Graph&lt;/strong&gt;; queries are not sent to third-party search engines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Admin-level domain filters&lt;/strong&gt; that agents can't bypass, plus per-call domain, published-date, and country filters.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;$10 per 1,000 tool calls&lt;/strong&gt;, plus $0.005 per 1,000 Gateway invocations. Web Search is $7 per 1,000 queries.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;us-east-1 only&lt;/strong&gt;, with a launch quota of &lt;strong&gt;10 TPS per account&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Terms require you to &lt;strong&gt;display the citations&lt;/strong&gt; with any result you show users and prohibit bulk storage of results.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fviagubgvtqt5helf9g64.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fviagubgvtqt5helf9g64.webp" alt="Request flow from an agent through AgentCore Gateway to Web Answer and back with citations" width="800" height="210"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Request flow: the Gateway calls Web Answer with its own service role and returns a cited answer. Image: AWS.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Web Search vs. Web Answer
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Web Search&lt;/th&gt;
&lt;th&gt;Web Answer&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Returns&lt;/td&gt;
&lt;td&gt;URLs, titles, snippets, dates&lt;/td&gt;
&lt;td&gt;Markdown answer with &lt;code&gt;[[N](url)]&lt;/code&gt; citations&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Who synthesizes&lt;/td&gt;
&lt;td&gt;Your model&lt;/td&gt;
&lt;td&gt;An AWS-managed agent&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tokens in your model&lt;/td&gt;
&lt;td&gt;Many&lt;/td&gt;
&lt;td&gt;Few&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Price&lt;/td&gt;
&lt;td&gt;$7 / 1,000 queries&lt;/td&gt;
&lt;td&gt;$10 / 1,000 calls&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best for&lt;/td&gt;
&lt;td&gt;Broad research, fine control&lt;/td&gt;
&lt;td&gt;Specific factual questions&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;AWS suggests using both. I agree: Web Search when your model needs to reason across many results, Web Answer when you just need a current, verifiable fact without burning your orchestrator's context window on ten web pages.&lt;/p&gt;

&lt;p&gt;One detail I like: the Gateway service role &lt;strong&gt;does not need &lt;code&gt;bedrock:InvokeModel&lt;/code&gt;&lt;/strong&gt;. The model Web Answer uses runs on AWS's side and is included in the per-call price. Processing may use cross-Region inference, but within the same geography (US, EU, or APAC).&lt;/p&gt;
&lt;h2&gt;
  
  
  The response format
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;content[0].text&lt;/code&gt; is ready-to-render markdown with citation markers; &lt;code&gt;structuredContent.answer[]&lt;/code&gt; holds blocks of &lt;code&gt;text&lt;/code&gt; plus a &lt;code&gt;citations&lt;/code&gt; list (&lt;code&gt;url&lt;/code&gt;, optional &lt;code&gt;domain&lt;/code&gt;). Marker &lt;code&gt;N&lt;/code&gt; maps to &lt;code&gt;citations[N-1]&lt;/code&gt; &lt;strong&gt;within the same block&lt;/strong&gt;, there can be several blocks, and on failure &lt;code&gt;isError&lt;/code&gt; is &lt;code&gt;true&lt;/code&gt; with the reason in &lt;code&gt;content[0].text&lt;/code&gt;. AWS's example shows a server-side &lt;code&gt;latencyMs&lt;/code&gt; of about 5.2 seconds.&lt;/p&gt;
&lt;h2&gt;
  
  
  The filter model is a platform team's friend
&lt;/h2&gt;

&lt;p&gt;Two levels, one rule: &lt;strong&gt;the admin sets the boundary, the agent can only narrow it&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Target level (admin):&lt;/strong&gt; &lt;code&gt;parameterValues.domainFilter&lt;/code&gt; with &lt;code&gt;include&lt;/code&gt; and &lt;code&gt;exclude&lt;/code&gt; lists of up to 50 domains each, set when you create the target.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Call level (agent):&lt;/strong&gt; &lt;code&gt;filters.domainFilter&lt;/code&gt;, &lt;code&gt;filters.publishedDateFilter&lt;/code&gt; (&lt;code&gt;from&lt;/code&gt;/&lt;code&gt;to&lt;/code&gt;, inclusive, UTC, ISO-8601) and &lt;code&gt;userLocation.country&lt;/code&gt; (ISO 3166-1 alpha-2).&lt;/li&gt;
&lt;li&gt;Excludes are merged (up to 100). If both levels set &lt;code&gt;include&lt;/code&gt;, the intersection is used, and an empty intersection is a validation error. Excluding a root domain also excludes its subdomains.&lt;/li&gt;
&lt;li&gt;Per-call filters and &lt;code&gt;userLocation&lt;/code&gt; require &lt;strong&gt;connector version 1.1.0+&lt;/strong&gt;. Queries can be up to 30,000 characters.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6l6xwatbgucv9gatjeex.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6l6xwatbgucv9gatjeex.webp" alt="How target-level and call-level domain filters are combined" width="800" height="291"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;How the two filter levels combine. Diagram: CloudAcademy.ar (Spanish).&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;

&lt;p&gt;If you don't have a Gateway yet, the console wizard creates one in a few steps: MCP target, &lt;strong&gt;Connectors&lt;/strong&gt; target type, then pick the tool. (The screenshot below is from the Web Search launch; Web Answer is added from the same place.)&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1cfk9512m5z8xa4dhbgo.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1cfk9512m5z8xa4dhbgo.webp" alt="AgentCore console: adding a Connectors target to a Gateway" width="800" height="876"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Adding a connector target in the AgentCore console. Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Let the Gateway service role invoke Web Answer.&lt;/strong&gt; Note the &lt;code&gt;aws&lt;/code&gt; account in the resource ARN: it's an AWS-owned tool.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"Version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"2012-10-17"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"Statement"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"Effect"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Allow"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"Action"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"bedrock-agentcore:InvokeWebAnswer"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"Resource"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"arn:aws:bedrock-agentcore:us-east-1:aws:tool/web-answer.v1"&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;2. Create the target with an allowlist.&lt;/strong&gt; This one turns Web Answer into an "official docs only" assistant without running your own RAG:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws bedrock-agentcore-control create-gateway-target \
  --gateway-identifier &lt;span class="s2"&gt;"&amp;lt;GATEWAY_ID&amp;gt;"&lt;/span&gt; \
  --name &lt;span class="s2"&gt;"web-answer-tool"&lt;/span&gt; \
  --target-configuration &lt;span class="s1"&gt;'{
    "mcp": {
      "connector": {
        "source": { "connectorId": "web-answer" },
        "configurations": [{
          "name": "WebAnswer",
          "parameterValues": {
            "domainFilter": {
              "include": ["docs.aws.amazon.com", "aws.amazon.com", "kubernetes.io"]
            }
          }
        }]
      }
    }
  }'&lt;/span&gt; \
  --credential-provider-configurations &lt;span class="s1"&gt;'[{"credentialProviderType": "GATEWAY_IAM_ROLE"}]'&lt;/span&gt; \
  --region us-east-1
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;3. Test it with MCP Inspector&lt;/strong&gt; before writing code. If semantic search is enabled on the Gateway, &lt;code&gt;tools/list&lt;/code&gt; also returns &lt;code&gt;x_amz_bedrock_agentcore_search&lt;/code&gt;, possibly first, so match on the &lt;code&gt;___WebAnswer&lt;/code&gt; suffix rather than position.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fty8d3of83work2w02gk2.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fty8d3of83work2w02gk2.webp" alt="MCP Inspector connected to an AgentCore Gateway running a tool" width="800" height="559"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Testing a Gateway tool with MCP Inspector (Web Search launch screenshot). Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Call it from Strands Agents&lt;/strong&gt; with SigV4, no API keys:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;mcp_proxy_for_aws&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;aws_iam_streamablehttp_client&lt;/span&gt;
&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;strands&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;Agent&lt;/span&gt;
&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;strands&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;models&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;BedrockModel&lt;/span&gt;
&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;strands&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;mcp&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;MCPClient&lt;/span&gt;

&lt;span class="no"&gt;GATEWAY_URL&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"https://&amp;lt;gateway-id&amp;gt;.gateway.bedrock-agentcore.us-east-1.amazonaws.com/mcp"&lt;/span&gt;

&lt;span class="n"&gt;gateway&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;MCPClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;lambda&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;aws_iam_streamablehttp_client&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;endpoint&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="no"&gt;GATEWAY_URL&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;aws_region&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"us-east-1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;aws_service&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"bedrock-agentcore"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;with&lt;/span&gt; &lt;span class="n"&gt;gateway&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;agent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nc"&gt;BedrockModel&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;model_id&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"us.anthropic.claude-sonnet-5"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;max_tokens&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;1024&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;tools&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;gateway&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;list_tools_sync&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;system_prompt&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"Use WebAnswer only for facts that change over time, and always show its sources."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"Which Kubernetes versions does Amazon EKS currently offer in standard support?"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;message&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;5. Running on EKS.&lt;/strong&gt; The pod's role (via EKS Pod Identity) needs &lt;code&gt;bedrock-agentcore:InvokeGateway&lt;/code&gt; on the Gateway ARN plus Bedrock permissions for your orchestrator model; it does not need Web Answer permissions, since the Gateway role handles that. For private clusters, the Gateway supports PrivateLink (&lt;code&gt;com.amazonaws.us-east-1.bedrock-agentcore.gateway&lt;/code&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;Web Answer is &lt;strong&gt;$10 per 1,000 tool calls&lt;/strong&gt;; Gateway invocations are $0.005 per 1,000. A back-of-the-envelope example: an internal assistant doing 2,000 questions per business day (about 44,000 a month) is roughly &lt;strong&gt;$440/month&lt;/strong&gt; in Web Answer, plus your orchestrator model, which should spend fewer tokens than with raw Web Search results. Keep costs in check by telling the agent to call it only for time-sensitive facts, and consider Web Search at $7 for long research loops where your model filters anyway.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to check before production
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Region:&lt;/strong&gt; us-east-1 only, more "coming soon". If your Gateway is in sa-east-1 or eu-west-1, you'll need one in us-east-1 for this tool, with the cross-Region latency and transfer that implies.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Data residency:&lt;/strong&gt; AWS says queries stay within AWS, aren't sent to third-party engines, and AWS's diagram says they aren't used to train models. Still, if your agent puts customer data into the &lt;code&gt;query&lt;/code&gt;, you're shipping it to a US Region. Guardrails or a Gateway interceptor help.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Quota:&lt;/strong&gt; 10 TPS per account at launch. A runaway agent loop or several teams in one account will hit it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Terms:&lt;/strong&gt; you must retain and display citations; no bulk extraction or building your own index from results.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Unknowns:&lt;/strong&gt; I couldn't find documented retention for queries or how robots.txt is handled in the index. Ask your account team if that matters to you.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;Web Answer turns web research into a managed primitive with clear separation of duties: the agent decides when it needs external information, the Gateway decides who can use it and with which sources, and AWS does the searching and citing. For platform teams, that's governance you previously had to build. The trade-off is a black box: you don't choose the model, you don't see intermediate searches, and you depend on the quality of Amazon's index. Start with a narrow, strict-allowlist use case, measure quality and cost for a few weeks, then widen it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/11/web-answer-amazon-bedrock/" rel="noopener noreferrer"&gt;What's New announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/machine-learning/getting-started-with-web-answer-on-amazon-bedrock-agentcore/" rel="noopener noreferrer"&gt;Getting started with Web Answer (AWS AI Blog)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-target-connector-web-answer-tool.html" rel="noopener noreferrer"&gt;Web Answer documentation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-add-target-api-target-config.html" rel="noopener noreferrer"&gt;Gateway target configuration&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/bedrock/agentcore/pricing/" rel="noopener noreferrer"&gt;AgentCore pricing&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://blog.cloudacademy.ar/web-answer-bedrock-agentcore-respuestas-con-citas-para-agentes/" rel="noopener noreferrer"&gt;Full write-up in Spanish on CloudAcademy.ar&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>ai</category>
      <category>agents</category>
      <category>mcp</category>
    </item>
    <item>
      <title>DynamoDB Bulk Put Items: Load S3 Data Into Live Tables Without Your Own Pipeline</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Tue, 10 Nov 2026 14:45:50 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/dynamodb-bulk-put-items-load-s3-data-into-live-tables-without-your-own-pipeline-lak</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/dynamodb-bulk-put-items-load-s3-data-into-live-tables-without-your-own-pipeline-lak</guid>
      <description>&lt;p&gt;If you have ever owned a nightly job that pushes millions of items into a live DynamoDB table, you know the drill: a Glue job or a container calling &lt;code&gt;BatchWriteItem&lt;/code&gt; 25 items at a time, retry logic for unprocessed items, a capacity bump before the window, alarms muted by hand, and a "report" that is mostly log grepping.&lt;/p&gt;

&lt;p&gt;On November 5, 2026, AWS made &lt;strong&gt;bulk put items for Amazon DynamoDB&lt;/strong&gt; generally available. It is a managed, asynchronous operation that reads files from Amazon S3 and writes them into an &lt;strong&gt;existing&lt;/strong&gt; table, including one that is serving production traffic, using dedicated capacity that does not touch your table's throughput. This post covers how it works, how to run it from the AWS CLI and from a Kubernetes CronJob on EKS, what it costs, and the gotchas worth knowing before you migrate a pipeline to it.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Loads data from S3 into an &lt;strong&gt;existing table&lt;/strong&gt;, in the background, on &lt;strong&gt;dedicated capacity&lt;/strong&gt;. Bulk writes do not show up in table consumed-capacity metrics, so your alarms and auto scaling keep seeing real application traffic only.&lt;/li&gt;
&lt;li&gt;Every input line can carry its own &lt;strong&gt;condition expression&lt;/strong&gt; (same grammar as &lt;code&gt;PutItem&lt;/code&gt;): insert-only writes, version-guarded writes, or unconditional ones, mixed in the same file.&lt;/li&gt;
&lt;li&gt;Writes a &lt;strong&gt;report to S3&lt;/strong&gt; with failed items and, optionally, condition-skipped items, including &lt;code&gt;newImage&lt;/code&gt; and &lt;code&gt;oldImage&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Two tiers: &lt;strong&gt;Standard&lt;/strong&gt; ($0.10/GB, up to 100 GB per operation) and &lt;strong&gt;Accelerated&lt;/strong&gt; ($0.20/GB, up to 1 TB), US East (N. Virginia) prices.&lt;/li&gt;
&lt;li&gt;Standard is in all commercial Regions and GovCloud (US); Accelerated is in 23 Regions, including South America (São Paulo).&lt;/li&gt;
&lt;li&gt;No global tables at launch, whole-item writes only, not transactional.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyxwi7utzdt5xqd4tubr8.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyxwi7utzdt5xqd4tubr8.webp" alt="Bulk put items architecture: S3 input, bulk operation, dedicated capacity pool, S3 report" width="800" height="328"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Bulk put items writes through a dedicated capacity pool and drops its report in S3. Image: AWS.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Where it fits
&lt;/h2&gt;

&lt;p&gt;DynamoDB already had two bulk paths, and the new one sits right between them:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;BatchWriteItem&lt;/th&gt;
&lt;th&gt;Import from S3&lt;/th&gt;
&lt;th&gt;Bulk put items&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Target table&lt;/td&gt;
&lt;td&gt;Existing&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;New&lt;/strong&gt; (it creates it)&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;Existing&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Who writes&lt;/td&gt;
&lt;td&gt;Your code&lt;/td&gt;
&lt;td&gt;DynamoDB&lt;/td&gt;
&lt;td&gt;DynamoDB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Capacity&lt;/td&gt;
&lt;td&gt;Your table's&lt;/td&gt;
&lt;td&gt;None consumed&lt;/td&gt;
&lt;td&gt;Dedicated, isolated&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Per-item conditions&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;td&gt;Yes, per line&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Error report&lt;/td&gt;
&lt;td&gt;Build your own&lt;/td&gt;
&lt;td&gt;CloudWatch Logs&lt;/td&gt;
&lt;td&gt;S3, with old/new images&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;What it is &lt;strong&gt;not&lt;/strong&gt;: it is not a partial update (use &lt;code&gt;UpdateItem&lt;/code&gt; for attribute-level changes), it is not transactional (cancelling leaves already-written items in place), and it does not support global tables yet.&lt;/p&gt;
&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Put line-delimited &lt;strong&gt;DynamoDB JSON&lt;/strong&gt; or &lt;strong&gt;Amazon Ion&lt;/strong&gt; files under an S3 prefix, optionally gzip or zstd compressed.&lt;/li&gt;
&lt;li&gt;Call &lt;code&gt;StartBulkPutItems&lt;/code&gt;. It returns immediately with an operation ARN.&lt;/li&gt;
&lt;li&gt;DynamoDB reads, validates, and writes in the background. GSIs and LSIs are updated (index replication also runs on isolated capacity), and items show up in &lt;strong&gt;DynamoDB Streams and Kinesis Data Streams&lt;/strong&gt; like any other write, so zero-ETL integrations and stream consumers will see them.&lt;/li&gt;
&lt;li&gt;The operation ends as &lt;code&gt;COMPLETED&lt;/code&gt;, &lt;code&gt;COMPLETED_WITH_ERRORS&lt;/code&gt;, &lt;code&gt;FAILED&lt;/code&gt;, or &lt;code&gt;CANCELLED&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Expect counters at zero for the first few minutes while the operation provisions, and note that they refresh periodically, not in real time.&lt;/p&gt;
&lt;h2&gt;
  
  
  Conditions per line are the real feature
&lt;/h2&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"Item"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"pk"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"S"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"SKU#1001"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"S"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Gourd mate"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"stock"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"N"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"40"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"Item"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"pk"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"S"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"SKU#1002"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"S"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Alpaca bombilla"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"ConditionExpression"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"attribute_not_exists(pk)"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"Item"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"pk"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"S"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"SKU#1003"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"price"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"N"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"8900"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"N"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"12"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"ConditionExpression"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"attribute_not_exists(#v) OR #v &amp;lt; :v"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"ExpressionAttributeNames"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"#v"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"version"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"ExpressionAttributeValues"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;":v"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"N"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"12"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;


&lt;p&gt;The classic trap: writing only &lt;code&gt;#v &amp;lt; :v&lt;/code&gt;. That condition &lt;strong&gt;fails when the item does not exist yet&lt;/strong&gt;, because there is nothing to compare. Classmethod hit exactly this in their day-one test. Use &lt;code&gt;attribute_not_exists(#v) OR #v &amp;lt; :v&lt;/code&gt;, as the docs show.&lt;/p&gt;

&lt;p&gt;Condition-skipped rows are not failures; they count as processed and land in the &lt;code&gt;CONDITION_CHECK_FAILURE&lt;/code&gt; report. They are, however, &lt;strong&gt;billed&lt;/strong&gt;, even against non-existent items (at the 1 KB minimum).&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fv44pxzrzngero54v66qe.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fv44pxzrzngero54v66qe.webp" alt="Lifecycle of an idempotent bulk put operation and how to use the reports" width="799" height="366"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Operation lifecycle and the retry loop. Diagram: CloudAcademy.ar (Spanish).&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;

&lt;p&gt;You need &lt;strong&gt;AWS CLI 2.39.4 or later&lt;/strong&gt;. Permissions: account-scoped &lt;code&gt;dynamodb:StartBulkOperation&lt;/code&gt;, table-scoped &lt;code&gt;dynamodb:BulkPutItems&lt;/code&gt;, read on the input bucket, write on the report bucket, and KMS access if you use customer managed keys. The operation runs with the caller's permissions.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws dynamodb start-bulk-put-items \
  --table-name catalog \
  --input-s3-configuration &lt;span class="s1"&gt;'{
    "S3Bucket": "my-input-bucket",
    "S3KeyPrefix": "catalog/2026-11-09/",
    "InputFormat": "DYNAMODB_JSON",
    "InputCompressionType": "GZIP"
  }'&lt;/span&gt; \
  --output-s3-configuration &lt;span class="s1"&gt;'{
    "S3Bucket": "my-report-bucket",
    "S3KeyPrefix": "bulk/catalog/",
    "OutputType": ["ERROR", "CONDITION_CHECK_FAILURE"]
  }'&lt;/span&gt; \
  --performance-tier STANDARD \
  --client-token &lt;span class="s2"&gt;"catalog-2026-11-09"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Always pass &lt;code&gt;--client-token&lt;/code&gt;: the same token and parameters within &lt;strong&gt;8 hours&lt;/strong&gt; return the same operation instead of starting a duplicate, which is exactly what you want when a scheduler retries. &lt;code&gt;--max-write-request-units&lt;/code&gt; (minimum 1,024) caps the write rate of this operation only; set it too low and the operation can hit its time limit and fail with &lt;code&gt;Timeout&lt;/code&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws dynamodb describe-bulk-operation \
  --bulk-operation-arn &lt;span class="s2"&gt;"$ARN"&lt;/span&gt; \
  --query &lt;span class="s1"&gt;'BulkOperationDescription.{status:BulkOperationStatus,processed:ProcessedItemCount,failed:FailedItemCount}'&lt;/span&gt;

aws dynamodb list-bulk-operations --max-items 25    &lt;span class="c"&gt;# 90 days of history&lt;/span&gt;
aws dynamodb cancel-bulk-operation --bulk-operation-arn &lt;span class="s2"&gt;"$ARN"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Running it from EKS
&lt;/h3&gt;

&lt;p&gt;If your platform already lives on EKS, you don't need Glue or Step Functions for this. A CronJob with the AWS CLI image and EKS Pod Identity is enough, and making &lt;code&gt;COMPLETED_WITH_ERRORS&lt;/code&gt; fail the Job surfaces partial loads in your existing Kubernetes alerting:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;apiVersion&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;batch/v1&lt;/span&gt;
&lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;CronJob&lt;/span&gt;
&lt;span class="na"&gt;metadata&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;catalog-bulk-put&lt;/span&gt;
  &lt;span class="na"&gt;namespace&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;data-jobs&lt;/span&gt;
&lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;schedule&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"0 4 * * *"&lt;/span&gt;
  &lt;span class="na"&gt;concurrencyPolicy&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Forbid&lt;/span&gt;
  &lt;span class="na"&gt;jobTemplate&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;backoffLimit&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;
      &lt;span class="na"&gt;template&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
          &lt;span class="na"&gt;serviceAccountName&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;catalog-bulk-put&lt;/span&gt;   &lt;span class="c1"&gt;# mapped to an IAM role via EKS Pod Identity
&lt;/span&gt;          &lt;span class="na"&gt;restartPolicy&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Never&lt;/span&gt;
          &lt;span class="na"&gt;containers&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;bulk-put&lt;/span&gt;
              &lt;span class="na"&gt;image&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;amazon/aws-cli:latest&lt;/span&gt;         &lt;span class="c1"&gt;# pin a version &amp;gt;= 2.39.4
&lt;/span&gt;              &lt;span class="na"&gt;command&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"/bin/bash"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"-c"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
              &lt;span class="na"&gt;args&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
                &lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="p"&gt;|&lt;/span&gt;
                  &lt;span class="s"&gt;set -euo pipefail&lt;/span&gt;
                  &lt;span class="s"&gt;BATCH=$(date -u +%F)&lt;/span&gt;
                  &lt;span class="s"&gt;ARN=$(aws dynamodb start-bulk-put-items \&lt;/span&gt;
                    &lt;span class="s"&gt;--table-name catalog \&lt;/span&gt;
                    &lt;span class="s"&gt;--input-s3-configuration "{\"S3Bucket\":\"my-input-bucket\",\"S3KeyPrefix\":\"catalog/${BATCH}/\",\"InputFormat\":\"DYNAMODB_JSON\",\"InputCompressionType\":\"GZIP\"}" \&lt;/span&gt;
                    &lt;span class="s"&gt;--output-s3-configuration "{\"S3Bucket\":\"my-report-bucket\",\"S3KeyPrefix\":\"bulk/catalog/${BATCH}/\",\"OutputType\":[\"ERROR\",\"CONDITION_CHECK_FAILURE\"]}" \&lt;/span&gt;
                    &lt;span class="s"&gt;--performance-tier STANDARD \&lt;/span&gt;
                    &lt;span class="s"&gt;--client-token "catalog-${BATCH}" \&lt;/span&gt;
                    &lt;span class="s"&gt;--query 'BulkOperationDescription.BulkOperationArn' --output text)&lt;/span&gt;
                  &lt;span class="s"&gt;while true; do&lt;/span&gt;
                    &lt;span class="s"&gt;STATUS=$(aws dynamodb describe-bulk-operation --bulk-operation-arn "$ARN" \&lt;/span&gt;
                      &lt;span class="s"&gt;--query 'BulkOperationDescription.BulkOperationStatus' --output text)&lt;/span&gt;
                    &lt;span class="s"&gt;echo "$(date -u +%T) ${STATUS}"&lt;/span&gt;
                    &lt;span class="s"&gt;case "$STATUS" in&lt;/span&gt;
                      &lt;span class="s"&gt;IN_PROGRESS|CANCELLING) sleep 60 ;;&lt;/span&gt;
                      &lt;span class="s"&gt;COMPLETED) exit 0 ;;&lt;/span&gt;
                      &lt;span class="nv"&gt;*)&lt;/span&gt; &lt;span class="s"&gt;exit 1 ;;&lt;/span&gt;
                    &lt;span class="s"&gt;esac&lt;/span&gt;
                  &lt;span class="s"&gt;done&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Gotchas
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;One bulk operation per table at a time, and not while an index is being created.&lt;/li&gt;
&lt;li&gt;Up to 50,000 input objects and 5 GB per object; 5 Standard plus 5 Accelerated operations in flight per account.&lt;/li&gt;
&lt;li&gt;The table must be &lt;code&gt;ACTIVE&lt;/code&gt;; restoring tables are not supported.&lt;/li&gt;
&lt;li&gt;Size your stream consumers for normal traffic &lt;strong&gt;plus&lt;/strong&gt; the import, or throttle the import with &lt;code&gt;--max-write-request-units&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Accelerated does not support tables with vector indexes.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;Billing is per GB written, with a &lt;strong&gt;1 KB minimum per item&lt;/strong&gt;, metered separately for the table and &lt;strong&gt;each GSI&lt;/strong&gt; the write propagates to. In us-east-1 that's &lt;strong&gt;$0.10/GB on Standard and $0.20/GB on Accelerated&lt;/strong&gt;. AWS's example: 100 million 1 KB items = 100 GB = $10 (Standard) or $20 (Accelerated). Validation-rejected items are free; failed condition checks are not. Reports are billed at normal S3 rates.&lt;/p&gt;

&lt;p&gt;Practical ways to keep it cheap: audit which GSIs project the full item, avoid the 1 KB floor with tiny items where you can, send deltas rather than full snapshots (conditions are a safety net, not a filter), and default to Standard. For Accelerated, AWS publishes estimates of under 10 minutes for 10M items, under an hour for 100M, and 2 to 4 hours for 1B, under ideal conditions; benchmark three to five runs before committing to a window.&lt;/p&gt;

&lt;h2&gt;
  
  
  Region notes
&lt;/h2&gt;

&lt;p&gt;Standard is available in every commercial Region and GovCloud (US). Accelerated covers 23 Regions, and &lt;strong&gt;sa-east-1 is one of them&lt;/strong&gt;, which is good news for teams in Latin America. Keep input and report buckets in the same Region as the table to avoid cross-Region transfer and data residency questions, and treat the report bucket as sensitive: it contains copies of your items.&lt;/p&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;The big win is not the per-GB price; it is that AWS decoupled bulk loads from table capacity, both in resources and in metrics. Your alarms go back to measuring real traffic, and the load becomes a first-class object with status, history, CloudTrail audit, and a retryable report. Per-line conditions make imports safe to re-run, which is what you need when a scheduler retries at 4 a.m. What's missing: global tables support and a native completion event (EventBridge would be nice) so you don't have to poll. Until then, a CronJob with &lt;code&gt;describe-bulk-operation&lt;/code&gt; does the job.&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/11/amazon-dynamodb-now-supports-bulk-put-items/" rel="noopener noreferrer"&gt;What's New announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/database/introducing-bulk-put-items-for-amazon-dynamodb-load-data-from-amazon-s3-into-existing-tables-at-scale/" rel="noopener noreferrer"&gt;AWS Database Blog: Introducing bulk put items&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/amazondynamodb/latest/developerguide/BulkPutItems.Requesting.html" rel="noopener noreferrer"&gt;Docs: running a bulk put items operation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/amazondynamodb/latest/developerguide/BulkPutItems.Tiers.html" rel="noopener noreferrer"&gt;Docs: performance tiers and Regions&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/goto/cli2/dynamodb-2012-08-10/StartBulkPutItems" rel="noopener noreferrer"&gt;CLI reference: start-bulk-put-items&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://blog.cloudacademy.ar/dynamodb-bulk-put-items-carga-masiva-desde-s3-en-tablas-existentes/" rel="noopener noreferrer"&gt;Full write-up in Spanish on CloudAcademy.ar&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>dynamodb</category>
      <category>devops</category>
      <category>kubernetes</category>
    </item>
    <item>
      <title>EKS Auto Mode Now Speaks Gateway API: Managed ALBs and NLBs from Gateway and HTTPRoute</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Tue, 03 Nov 2026 12:09:51 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/eks-auto-mode-now-speaks-gateway-api-managed-albs-and-nlbs-from-gateway-and-httproute-7pl</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/eks-auto-mode-now-speaks-gateway-api-managed-albs-and-nlbs-from-gateway-and-httproute-7pl</guid>
      <description>&lt;p&gt;On October 28, 2026, Amazon EKS Auto Mode gained native support for the Kubernetes Gateway API. You define &lt;code&gt;GatewayClass&lt;/code&gt;, &lt;code&gt;Gateway&lt;/code&gt; and &lt;code&gt;HTTPRoute&lt;/code&gt; (or &lt;code&gt;GRPCRoute&lt;/code&gt;, &lt;code&gt;TCPRoute&lt;/code&gt;, &lt;code&gt;UDPRoute&lt;/code&gt;), and Auto Mode provisions and runs the Application Load Balancer or Network Load Balancer behind them. No AWS Load Balancer Controller to install, no Helm chart to upgrade, no IAM role to wire up.&lt;/p&gt;

&lt;p&gt;If your clusters already run Auto Mode, this closes an awkward gap. Auto Mode promised that AWS operates the cluster's infrastructure components, but its managed load balancing only spoke Ingress and &lt;code&gt;Service&lt;/code&gt; type &lt;code&gt;LoadBalancer&lt;/code&gt;. With Ingress NGINX retired in March and most teams moving to Gateway API, Auto Mode users had to bolt a controller back on, which is exactly what Auto Mode is supposed to remove.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmh4r7ihp34bo2azx0om9.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmh4r7ihp34bo2azx0om9.webp" alt="Who defines what with Gateway API on EKS Auto Mode" width="799" height="304"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Platform defines the class, cluster operators define the Gateway, app teams define routes, and Auto Mode runs the load balancer (labels in Spanish). Diagram: CloudAcademy.ar.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Two AWS-owned controllers: &lt;code&gt;eks.amazonaws.com/gateway-alb&lt;/code&gt; (L7, ALB) and &lt;code&gt;eks.amazonaws.com/gateway-nlb&lt;/code&gt; (L4, NLB). They coexist with Envoy Gateway, Istio and others.&lt;/li&gt;
&lt;li&gt;HTTP and gRPC routing, header matching, weighted traffic splitting and cross-namespace routes.&lt;/li&gt;
&lt;li&gt;ACM certificates are discovered automatically from the listener &lt;code&gt;hostname&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;AWS-specific settings live in three extension CRDs: &lt;code&gt;LoadBalancerConfiguration&lt;/code&gt;, &lt;code&gt;TargetGroupConfiguration&lt;/code&gt; and &lt;code&gt;ListenerRuleConfiguration&lt;/code&gt; (including Cognito/OIDC auth).&lt;/li&gt;
&lt;li&gt;Ingress and &lt;code&gt;LoadBalancer&lt;/code&gt; Services keep working, so you can migrate one app at a time.&lt;/li&gt;
&lt;li&gt;No extra charge beyond Auto Mode and the load balancers it creates.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;Gateway API splits responsibility into three layers, which is its main advantage over Ingress: the platform team owns &lt;code&gt;GatewayClass&lt;/code&gt;, cluster operators own &lt;code&gt;Gateway&lt;/code&gt; (listeners, hostnames, which namespaces may attach routes), and application teams own routes in their own namespaces.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fq24ogr0ddolphh2we4ou.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fq24ogr0ddolphh2we4ou.webp" alt="Gateway API roles" width="568" height="496"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Gateway API's role model, as presented in AWS's Load Balancer Controller post (adapted from the Gateway API project). Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Auto Mode watches those resources, maps them to load balancers, listeners, rules and target groups, corrects drift, and writes status (including the load balancer address) back to the Gateway and routes. It's the same reconciliation logic as the AWS Load Balancer Controller, just operated by AWS.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuck7miyls6f6xs2d2kph.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fuck7miyls6f6xs2d2kph.webp" alt="Gateway API controller reconciliation flow" width="768" height="576"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Controller reconciliation flow. Image: AWS.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Enabling it
&lt;/h2&gt;

&lt;p&gt;It's a cluster-level setting under &lt;code&gt;kubernetesNetworkConfig.elasticLoadBalancing.gatewayApi&lt;/code&gt;. With &lt;code&gt;crds.managed: true&lt;/code&gt;, AWS installs and maintains the Gateway API v1 CRDs (and refuses to touch existing CRDs older than v1). With &lt;code&gt;false&lt;/code&gt;, you own the CRDs, which is what you want if another implementation needs a specific version.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws eks update-cluster-config &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--name&lt;/span&gt; &lt;span class="nv"&gt;$CLUSTER_NAME&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--kubernetes-network-config&lt;/span&gt; &lt;span class="s1"&gt;'{"elasticLoadBalancing":{"enabled":true,"gatewayApi":{"crds":{"managed":true}}}}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Don't forget subnet tags. Auto Mode uses them to place public and internal load balancers, and missing tags are the number one reason a Gateway never becomes &lt;code&gt;Programmed&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  A realistic example
&lt;/h2&gt;

&lt;p&gt;One shared Gateway in an &lt;code&gt;infra&lt;/code&gt; namespace, HTTP and HTTPS listeners, and only namespaces labeled &lt;code&gt;gateway-access=true&lt;/code&gt; can attach routes:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;apiVersion&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;gateway.networking.k8s.io/v1&lt;/span&gt;
&lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;GatewayClass&lt;/span&gt;
&lt;span class="na"&gt;metadata&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;eks-auto-alb&lt;/span&gt;
&lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;controllerName&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;eks.amazonaws.com/gateway-alb&lt;/span&gt;
&lt;span class="nn"&gt;---&lt;/span&gt;
&lt;span class="na"&gt;apiVersion&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;gateway.k8s.aws/v1&lt;/span&gt;
&lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;LoadBalancerConfiguration&lt;/span&gt;
&lt;span class="na"&gt;metadata&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;edge-public&lt;/span&gt;
  &lt;span class="na"&gt;namespace&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;infra&lt;/span&gt;
&lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;scheme&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;internet-facing&lt;/span&gt;
&lt;span class="nn"&gt;---&lt;/span&gt;
&lt;span class="na"&gt;apiVersion&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;gateway.networking.k8s.io/v1&lt;/span&gt;
&lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Gateway&lt;/span&gt;
&lt;span class="na"&gt;metadata&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;edge&lt;/span&gt;
  &lt;span class="na"&gt;namespace&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;infra&lt;/span&gt;
&lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;gatewayClassName&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;eks-auto-alb&lt;/span&gt;
  &lt;span class="na"&gt;infrastructure&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;parametersRef&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;group&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;gateway.k8s.aws&lt;/span&gt;
      &lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;LoadBalancerConfiguration&lt;/span&gt;
      &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;edge-public&lt;/span&gt;
  &lt;span class="na"&gt;listeners&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;https&lt;/span&gt;
      &lt;span class="na"&gt;protocol&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;HTTPS&lt;/span&gt;
      &lt;span class="na"&gt;port&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;443&lt;/span&gt;
      &lt;span class="na"&gt;hostname&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;api.example.com"&lt;/span&gt;
      &lt;span class="na"&gt;allowedRoutes&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
        &lt;span class="na"&gt;namespaces&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
          &lt;span class="na"&gt;from&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Selector&lt;/span&gt;
          &lt;span class="na"&gt;selector&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
            &lt;span class="na"&gt;matchLabels&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
              &lt;span class="na"&gt;gateway-access&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;true"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No certificate reference on the HTTPS listener: the controller finds a matching ACM certificate from the hostname and follows ACM updates, so renewals need no changes. Create the &lt;code&gt;GatewayClass&lt;/code&gt; first; there's a short delay before a new class can be referenced.&lt;/p&gt;

&lt;p&gt;The payments team then ships a 10% canary from its own namespace:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;apiVersion&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;gateway.networking.k8s.io/v1&lt;/span&gt;
&lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;HTTPRoute&lt;/span&gt;
&lt;span class="na"&gt;metadata&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;payments&lt;/span&gt;
  &lt;span class="na"&gt;namespace&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;payments&lt;/span&gt;
&lt;span class="na"&gt;spec&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;parentRefs&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;edge&lt;/span&gt;
      &lt;span class="na"&gt;namespace&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;infra&lt;/span&gt;
      &lt;span class="na"&gt;sectionName&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;https&lt;/span&gt;
  &lt;span class="na"&gt;rules&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;matches&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
        &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;path&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
            &lt;span class="na"&gt;type&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;PathPrefix&lt;/span&gt;
            &lt;span class="na"&gt;value&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;/payments&lt;/span&gt;
      &lt;span class="na"&gt;backendRefs&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
        &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;payments-v1&lt;/span&gt;
          &lt;span class="na"&gt;port&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;80&lt;/span&gt;
          &lt;span class="na"&gt;weight&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;90&lt;/span&gt;
        &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;payments-v2&lt;/span&gt;
          &lt;span class="na"&gt;port&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;80&lt;/span&gt;
          &lt;span class="na"&gt;weight&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;10&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Health checks and deregistration delay go in a &lt;code&gt;TargetGroupConfiguration&lt;/code&gt; attached to the Service (&lt;code&gt;spec.targetReference.name&lt;/code&gt;). Dropping the deregistration delay from the default 300 seconds to 30 makes rollouts noticeably faster.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;kubectl &lt;span class="nb"&gt;wait&lt;/span&gt; &lt;span class="nt"&gt;--for&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nv"&gt;condition&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;Programmed gateway/edge &lt;span class="nt"&gt;-n&lt;/span&gt; infra &lt;span class="nt"&gt;--timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;10m
kubectl get httproute &lt;span class="nt"&gt;-A&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If a route doesn't show up, check &lt;code&gt;status.parents&lt;/code&gt;: the controller records &lt;code&gt;Accepted&lt;/code&gt; and the reason when it fails.&lt;/p&gt;

&lt;h2&gt;
  
  
  What an ALB can't do
&lt;/h2&gt;

&lt;p&gt;This is the part to read before committing. Per the Load Balancer Controller compatibility table, &lt;code&gt;HTTPRoute&lt;/code&gt; doesn't support response header modification, request mirroring, CORS, timeouts, retries or default gateways; request header modification is partial; session persistence and external auth go through &lt;code&gt;ListenerRuleConfiguration&lt;/code&gt;; there's no &lt;code&gt;BackendTLSPolicy&lt;/code&gt;; and standard listener &lt;code&gt;certificateRefs&lt;/code&gt; aren't used. Redirects with &lt;code&gt;ReplacePrefixMatch&lt;/code&gt; must also change scheme, port or host, or the ALB rejects them as loops. If you depend on per-route timeouts and retries, you'll need a real L7 proxy for that part.&lt;/p&gt;

&lt;h2&gt;
  
  
  Migrating from Ingress
&lt;/h2&gt;

&lt;p&gt;Ingress and Gateway API run side by side, so migrate per app. AWS's &lt;code&gt;lbc-migrate&lt;/code&gt; toolkit (July 2026) translates Load Balancer Controller Ingress resources, annotations and IngressGroups into Gateway API manifests and supports dry-run Gateways. It targets the controller, so swap the &lt;code&gt;GatewayClass&lt;/code&gt; to &lt;code&gt;eks.amazonaws.com/gateway-alb&lt;/code&gt; afterwards. Then apply, verify, shift DNS (Route 53 weights help) and delete the old Ingress.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;Gateway API support has no fee of its own. You pay the EKS cluster ($0.10/hour on standard support), the Auto Mode management fee on the instances it runs, and the load balancers: in us-east-1 an ALB is $0.0225/hour plus $0.008 per LCU, an NLB $0.0225/hour plus $0.006 per NLCU.&lt;/p&gt;

&lt;p&gt;My own math: one shared Gateway is one ALB, about $16.40/month in hours before LCUs. Consolidating five standalone Ingress ALBs into one Gateway with per-namespace routes saves around $65/month in hours alone. Keep an eye on rule count, though: ALB rule evaluations beyond the first 10 consume LCUs.&lt;/p&gt;

&lt;h2&gt;
  
  
  Availability
&lt;/h2&gt;

&lt;p&gt;It's available in every Region where EKS Auto Mode is, including South America (São Paulo). The data plane is an ALB or NLB in your VPC and Region, so TLS terminates in-region. Cross-AZ traffic from the load balancer to pods is billed as usual.&lt;/p&gt;

&lt;h2&gt;
  
  
  Gotchas worth knowing before day one
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Old CRDs block managed mode.&lt;/strong&gt; If an earlier experiment left Gateway API CRDs below v1 in the cluster, the update is rejected and nothing changes. Upgrade them or switch to unmanaged.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;L4 routes are still experimental upstream.&lt;/strong&gt; &lt;code&gt;TCPRoute&lt;/code&gt; and &lt;code&gt;UDPRoute&lt;/code&gt; live in Gateway API's experimental channel, so confirm the CRDs you have installed include them before you rely on the NLB class.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Porting manifests from other implementations.&lt;/strong&gt; Routes written for Envoy Gateway or Istio often reference TLS Secrets, timeouts or retry filters. Those need rework: certificates come from ACM, and timeouts and retries aren't available on an ALB.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debug from status, not logs.&lt;/strong&gt; You don't have controller pods to tail. The &lt;code&gt;status&lt;/code&gt; of the Gateway and each route is your main debugging tool, so make &lt;code&gt;kubectl describe&lt;/code&gt; part of your runbooks.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Shared Gateway, shared blast radius.&lt;/strong&gt; A shared ALB is cheaper, but a bad rule in one route can affect lower-precedence rules in the same route. Keep routes small and owned by one team each.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;The most important part of this launch is about ownership, not features. With Ingress, the platform team either became the bottleneck for every routing change or handed out annotations that could break a shared load balancer. Gateway API fixes that with a clear permission model, and now AWS runs the implementation. The real constraint is that you get exactly what an ALB can do, and the unsupported list isn't short. For most APIs and frontends I see, it's plenty. Enable it in managed mode, migrate one boring app first, and use weighted canaries from day one.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Resources&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-eks-auto-mode-kubernetes-gateway-api/" rel="noopener noreferrer"&gt;What's New announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/eks/latest/userguide/auto-configure-gateway-api.html" rel="noopener noreferrer"&gt;EKS docs: Gateway API in Auto Mode&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://kubernetes-sigs.github.io/aws-load-balancer-controller/v3.5/guide/gateway/l7gateway/" rel="noopener noreferrer"&gt;ALB Gateway API guide and compatibility table&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/networking-and-content-delivery/aws-load-balancer-controller-adds-general-availability-support-for-kubernetes-gateway-api" rel="noopener noreferrer"&gt;AWS post: Load Balancer Controller Gateway API GA&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/networking-and-content-delivery/introducing-the-lbc-ingress-to-gateway-api-migration-toolkit/" rel="noopener noreferrer"&gt;AWS post: Ingress-to-Gateway API migration toolkit&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://aws.amazon.com/eks/pricing/" rel="noopener noreferrer"&gt;EKS pricing&lt;/a&gt; and &lt;a href="https://aws.amazon.com/elasticloadbalancing/pricing/" rel="noopener noreferrer"&gt;ELB pricing&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Originally published in Spanish at &lt;a href="https://blog.cloudacademy.ar/eks-auto-mode-gateway-api-alb-nlb-administrados/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>aws</category>
      <category>kubernetes</category>
      <category>eks</category>
      <category>devops</category>
    </item>
    <item>
      <title>Step Functions as AI Agent Tools: The Model Decides When, the Workflow Decides How</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Tue, 03 Nov 2026 12:09:47 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/step-functions-as-ai-agent-tools-the-model-decides-when-the-workflow-decides-how-2l8p</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/step-functions-as-ai-agent-tools-the-model-decides-when-the-workflow-decides-how-2l8p</guid>
      <description>&lt;p&gt;On October 29, 2026, AWS made it possible to expose existing AWS Step Functions workflows as tools for AI agents through Amazon Bedrock AgentCore Gateway. It sounds like just another integration, but it answers a question every team putting agents into production eventually hits: how much of a business process should the model be allowed to decide?&lt;/p&gt;

&lt;p&gt;AWS's answer is a clean split. The agent decides &lt;strong&gt;when&lt;/strong&gt; to start a process and &lt;strong&gt;with what input&lt;/strong&gt;. How that process runs (step order, retries, approvals, timeouts) stays in a Step Functions definition you wrote, versioned and tested. The model cannot skip the manager approval step because it "seemed reasonable".&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffgfz2fdrlkcxjs0ua0t3.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffgfz2fdrlkcxjs0ua0t3.webp" alt="How a call flows through the Step Functions connector" width="800" height="338"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Who decides what when an agent calls a workflow through the connector (labels in Spanish). Diagram: CloudAcademy.ar.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;AgentCore Gateway gets a managed connector (&lt;code&gt;connectorId: "states"&lt;/code&gt;) that turns Step Functions API operations into MCP tools. No glue Lambda, no custom MCP server.&lt;/li&gt;
&lt;li&gt;Ten operations are available: &lt;code&gt;StartExecution&lt;/code&gt;, &lt;code&gt;StartSyncExecution&lt;/code&gt;, &lt;code&gt;DescribeStateMachine&lt;/code&gt;, &lt;code&gt;ListExecutions&lt;/code&gt;, &lt;code&gt;DescribeExecution&lt;/code&gt;, &lt;code&gt;GetExecutionHistory&lt;/code&gt;, &lt;code&gt;StopExecution&lt;/code&gt;, and the three &lt;code&gt;SendTask*&lt;/code&gt; calls.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pinning&lt;/strong&gt; fixes the &lt;code&gt;stateMachineArn&lt;/code&gt; in the target configuration. The agent never sees it and cannot override it.&lt;/li&gt;
&lt;li&gt;Both Standard and Express workflows work, but history, describe, stop and task-token callbacks are Standard-only.&lt;/li&gt;
&lt;li&gt;All targets on a gateway share one service role. Different IAM boundaries mean different gateways.&lt;/li&gt;
&lt;li&gt;Available wherever AgentCore Gateway is, including São Paulo. No extra charge beyond standard Step Functions and Gateway pricing.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  What the connector actually is
&lt;/h2&gt;

&lt;p&gt;It is a managed target type for AgentCore Gateway. You create a target with &lt;code&gt;connectorId: "states"&lt;/code&gt;, choose which Step Functions operations become tools, and the gateway publishes them over MCP with a narrow input schema. Tool names are prefixed with the target name, for example &lt;code&gt;refunds___StartExecution&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;When the agent calls a tool, the gateway authenticates the caller (IAM/SigV4 or JWT), adds the pinned parameters, and calls Step Functions with its own service role, &lt;strong&gt;in its own Region&lt;/strong&gt;. A state machine in another Region is unreachable.&lt;/p&gt;

&lt;p&gt;What it is not: it is not an agent inside Step Functions (that is the separate AgentCore integration &lt;em&gt;from&lt;/em&gt; a workflow), it does not let the model design or change your workflow, and it is not business authorization. Being allowed to call &lt;code&gt;StartExecution&lt;/code&gt; on the refund workflow says nothing about whether &lt;em&gt;this&lt;/em&gt; refund makes sense. That still belongs inside the workflow.&lt;/p&gt;
&lt;h2&gt;
  
  
  Operations and pinning
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;StartExecution&lt;/code&gt;, &lt;code&gt;StartSyncExecution&lt;/code&gt; (Express only), &lt;code&gt;ListExecutions&lt;/code&gt; (Standard only) and &lt;code&gt;DescribeStateMachine&lt;/code&gt; accept a pin. The rest act on an execution ARN or a task token that the agent has to provide.&lt;/p&gt;

&lt;p&gt;By default &lt;code&gt;StartExecution&lt;/code&gt; exposes only &lt;code&gt;input&lt;/code&gt;. You can opt into &lt;code&gt;name&lt;/code&gt; and &lt;code&gt;traceHeader&lt;/code&gt; with &lt;code&gt;parameterOverrides&lt;/code&gt;. &lt;code&gt;stateMachineArn&lt;/code&gt; can never be exposed. If the agent sends one anyway, the call is rejected with &lt;code&gt;cannot set parameter(s): [/stateMachineArn]&lt;/code&gt; and nothing runs.&lt;/p&gt;

&lt;p&gt;One documented gotcha worth repeating: if you expose a pinnable operation and forget the pin, the target is created successfully and the tool fails on every call. The console doesn't show pinned values either, so verify with &lt;code&gt;get-gateway-target&lt;/code&gt;.&lt;/p&gt;
&lt;h2&gt;
  
  
  Standard vs. Express matters
&lt;/h2&gt;

&lt;p&gt;With &lt;strong&gt;Standard&lt;/strong&gt;, the agent can start, list, describe, read history, stop, and complete callback steps. Executions can last up to a year. With &lt;strong&gt;Express&lt;/strong&gt;, you get &lt;code&gt;StartSyncExecution&lt;/code&gt;, which returns the result in the same call, but &lt;code&gt;DescribeExecution&lt;/code&gt;, &lt;code&gt;GetExecutionHistory&lt;/code&gt; and &lt;code&gt;StopExecution&lt;/code&gt; return &lt;code&gt;InvalidArn&lt;/code&gt; on Express executions, and &lt;code&gt;.waitForTaskToken&lt;/code&gt; isn't supported. If your agent needs to "start it and check back later", use Standard.&lt;/p&gt;
&lt;h2&gt;
  
  
  Humans in the loop, and the catch
&lt;/h2&gt;

&lt;p&gt;The callback pattern is the most interesting part. A step with &lt;code&gt;.waitForTaskToken&lt;/code&gt; sends a token somewhere (SQS, a ticket, an approval UI) and waits until someone calls &lt;code&gt;SendTaskSuccess&lt;/code&gt; or &lt;code&gt;SendTaskFailure&lt;/code&gt;. A support agent can start a refund; above a threshold, the workflow parks until a human approves. The agent can only tell the customer "it's under review".&lt;/p&gt;

&lt;p&gt;The catch: IAM has &lt;strong&gt;no resource-level permissions&lt;/strong&gt; for &lt;code&gt;SendTaskSuccess&lt;/code&gt;, &lt;code&gt;SendTaskFailure&lt;/code&gt; and &lt;code&gt;SendTaskHeartbeat&lt;/code&gt;. They are granted on &lt;code&gt;"*"&lt;/code&gt;, so any agent holding a valid token can complete that task, even for state machines that aren't exposed on the gateway (an AWS Community Builder confirmed this in testing). Treat tokens as secrets and put &lt;code&gt;SendTask*&lt;/code&gt; tools on a separate gateway used only by whoever is supposed to approve.&lt;/p&gt;
&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;

&lt;p&gt;Everything below follows the official tutorial, adapted to a refund workflow. Use the same Region for all resources and a recent AWS CLI v2 (in community testing, 2.34.24 rejected the &lt;code&gt;connector&lt;/code&gt; block and 2.39.3 worked).&lt;/p&gt;

&lt;p&gt;The workflow keeps the decisions in ASL, not in the prompt:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"StartAt"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"CheckEligibility"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"States"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"CheckEligibility"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Task"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Resource"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"arn:aws:states:::lambda:invoke"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Parameters"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"FunctionName"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"check-refund"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"Payload.$"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"$"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"OutputPath"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"$.Payload"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Retry"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"ErrorEquals"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"States.TaskFailed"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"MaxAttempts"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"BackoffRate"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}],&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Next"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"NeedsApproval"&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"NeedsApproval"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Choice"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Choices"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"Variable"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"$.amount"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"NumericGreaterThan"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;500&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"Next"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"WaitForApproval"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}],&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Default"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"IssueCredit"&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"WaitForApproval"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Task"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Resource"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"arn:aws:states:::sqs:sendMessage.waitForTaskToken"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Parameters"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"QueueUrl"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://sqs.us-east-1.amazonaws.com/123456789012/approvals"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"MessageBody"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"order.$"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"$"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"taskToken.$"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"$$.Task.Token"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"TimeoutSeconds"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;86400&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Next"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"IssueCredit"&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"IssueCredit"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Task"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Resource"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"arn:aws:states:::lambda:invoke"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Parameters"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"FunctionName"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"issue-credit"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"Payload.$"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"$"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"OutputPath"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"$.Payload"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"End"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The gateway's service role trusts &lt;code&gt;bedrock-agentcore.amazonaws.com&lt;/code&gt; and needs two statements, because execution ARNs are a different resource type than state machine ARNs: &lt;code&gt;StartExecution&lt;/code&gt;/&lt;code&gt;ListExecutions&lt;/code&gt;/&lt;code&gt;DescribeStateMachine&lt;/code&gt; on &lt;code&gt;arn:aws:states:REGION:ACCOUNT:stateMachine:Refunds&lt;/code&gt;, and &lt;code&gt;DescribeExecution&lt;/code&gt;/&lt;code&gt;GetExecutionHistory&lt;/code&gt;/&lt;code&gt;StopExecution&lt;/code&gt; on &lt;code&gt;arn:aws:states:REGION:ACCOUNT:execution:Refunds:*&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Create the gateway and the connector target:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws bedrock-agentcore-control create-gateway &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--name&lt;/span&gt; &lt;span class="s2"&gt;"step-functions-gateway"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--protocol-type&lt;/span&gt; MCP &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--authorizer-type&lt;/span&gt; AWS_IAM &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--role-arn&lt;/span&gt; &lt;span class="s2"&gt;"arn:aws:iam::&lt;/span&gt;&lt;span class="k"&gt;${&lt;/span&gt;&lt;span class="nv"&gt;ACCOUNT_ID&lt;/span&gt;&lt;span class="k"&gt;}&lt;/span&gt;&lt;span class="s2"&gt;:role/AgentCoreGatewayToStepFunctions"&lt;/span&gt;

&lt;span class="nv"&gt;SM_ARN&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"arn:aws:states:&lt;/span&gt;&lt;span class="k"&gt;${&lt;/span&gt;&lt;span class="nv"&gt;AWS_REGION&lt;/span&gt;&lt;span class="k"&gt;}&lt;/span&gt;&lt;span class="s2"&gt;:&lt;/span&gt;&lt;span class="k"&gt;${&lt;/span&gt;&lt;span class="nv"&gt;ACCOUNT_ID&lt;/span&gt;&lt;span class="k"&gt;}&lt;/span&gt;&lt;span class="s2"&gt;:stateMachine:Refunds"&lt;/span&gt;

aws bedrock-agentcore-control create-gateway-target &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--gateway-identifier&lt;/span&gt; &lt;span class="s2"&gt;"GATEWAY_ID"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--name&lt;/span&gt; &lt;span class="s2"&gt;"refunds"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--target-configuration&lt;/span&gt; &lt;span class="s1"&gt;'{
    "mcp": { "connector": {
      "source": { "connectorId": "states" },
      "configurations": [
        { "name": "StartExecution", "parameterValues": { "stateMachineArn": "'&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$SM_ARN&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s1"&gt;'" } },
        { "name": "ListExecutions", "parameterValues": { "stateMachineArn": "'&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$SM_ARN&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s1"&gt;'" } },
        { "name": "DescribeExecution" },
        { "name": "GetExecutionHistory" }
      ]
    } }
  }'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--credential-provider-configurations&lt;/span&gt; &lt;span class="s1"&gt;'[{"credentialProviderType": "GATEWAY_IAM_ROLE"}]'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Call it the way an agent would, signing with SigV4 via &lt;code&gt;awscurl&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;awscurl &lt;span class="nt"&gt;--service&lt;/span&gt; bedrock-agentcore &lt;span class="nt"&gt;--region&lt;/span&gt; &lt;span class="nv"&gt;$AWS_REGION&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-X&lt;/span&gt; POST &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$GATEWAY_URL&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{
        "name":"refunds___StartExecution",
        "arguments":{"input":"{\"order\":\"A-1029\",\"amount\":720}"}}}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Note that &lt;code&gt;input&lt;/code&gt; is a &lt;strong&gt;string containing JSON&lt;/strong&gt;, not an object. It's the most common failure when a model builds the arguments.&lt;/p&gt;

&lt;p&gt;If your agents run on Amazon EKS, give the agent's service account a Pod Identity role with only &lt;code&gt;bedrock-agentcore:InvokeGateway&lt;/code&gt; on that gateway. The pod can reach one gateway; the gateway can reach specific state machines.&lt;/p&gt;

&lt;p&gt;A few more things the community found: &lt;code&gt;tools/list&lt;/code&gt; returns tools alphabetically; CloudTrail shows the gateway role with a fixed session name as the caller and hides &lt;code&gt;input&lt;/code&gt;, so you need your own correlation ID (inside &lt;code&gt;input&lt;/code&gt; or via &lt;code&gt;traceHeader&lt;/code&gt;) to tie an execution back to an agent session.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reference architectures from AWS
&lt;/h2&gt;

&lt;p&gt;AWS's Architecture Blog published the complementary pattern a week earlier: a workflow that owns every routing decision while an AgentCore agent reasons inside one bounded step.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpiw4utd6gh6bu5ochc17.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpiw4utd6gh6bu5ochc17.webp" alt="Medical coding workflow architecture" width="800" height="793"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;The workflow controls routing; the agent works inside a single step. Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;And MoEngage's write-up shows where this goes at scale: around 400 internal APIs exposed as agent tools through one gateway, with Cedar policies per agent.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx4z0nlugicyyy8ldbs05.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fx4z0nlugicyyy8ldbs05.webp" alt="MoEngage AgentCore Gateway architecture" width="800" height="375"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Gateway as the single tool catalog for an organization. Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;There's no connector fee. In us-east-1, Step Functions Standard costs $0.000025 per state transition (4,000 free per month), Express costs $1.00 per million requests plus $0.00001667 per GB-second, and AgentCore Gateway charges $0.005 per 1,000 invocations, $0.025 per 1,000 search calls and $0.02 per 100 indexed tools per month.&lt;/p&gt;

&lt;p&gt;My own rough math: 100,000 refunds a month on a six-transition Standard workflow is about $14.90 in transitions; five gateway calls per refund adds $2.50. Model tokens will dwarf both. The cheapest optimization is reducing polling: every &lt;code&gt;DescribeExecution&lt;/code&gt; is a gateway call &lt;em&gt;and&lt;/em&gt; another model turn, so for long processes, notify at the end instead.&lt;/p&gt;

&lt;h2&gt;
  
  
  Availability and data residency
&lt;/h2&gt;

&lt;p&gt;AgentCore Gateway, and therefore the connector, is available in 22 Regions including South America (São Paulo). Remember the gateway only calls Step Functions in its own Region, the agent's &lt;code&gt;input&lt;/code&gt; ends up in execution history (retention and encryption apply), and &lt;code&gt;includedData: "ALL_DATA"&lt;/code&gt; sends execution input and output back into the model's context.&lt;/p&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;The connector is useful, but the real value is the division of responsibility it forces: the model picks when and with what data, the workflow owns the how, and IAM plus pinning define the reach. It's the same rule we apply to humans in operations (you can request a deploy; the pipeline decides the steps and gates), applied to agents. What's missing is native correlation between an execution and the agent session that started it. Start with read-only or low-impact workflows, one gateway per permission boundary, and keep &lt;code&gt;SendTask*&lt;/code&gt; away from agents until you know exactly who approves what.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Resources&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/10/step-functions-workflows-ai-agent-tools/" rel="noopener noreferrer"&gt;What's New announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/step-functions/latest/dg/tutorial-agentcore-gateway.html" rel="noopener noreferrer"&gt;Official tutorial&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-target-connector-step-functions.html" rel="noopener noreferrer"&gt;Step Functions connector target reference&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://aws.amazon.com/bedrock/agentcore/pricing/" rel="noopener noreferrer"&gt;AgentCore pricing&lt;/a&gt; and &lt;a href="https://aws.amazon.com/step-functions/pricing/" rel="noopener noreferrer"&gt;Step Functions pricing&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;&lt;a href="https://dev.to/aws-builders/aws-using-step-functions-as-an-agent-tool-testing-how-pins-iam-and-tokens-work-agentcore-5a1o"&gt;Community test of pins, IAM and tokens&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/architecture/automate-medical-coding-with-step-functions-and-amazon-bedrock-agentcore/" rel="noopener noreferrer"&gt;Medical coding with Step Functions and AgentCore (AWS Architecture Blog)&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Originally published in Spanish at &lt;a href="https://blog.cloudacademy.ar/step-functions-como-herramientas-de-agentes-agentcore-gateway/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>aws</category>
      <category>ai</category>
      <category>serverless</category>
      <category>stepfunctions</category>
    </item>
    <item>
      <title>Structured outputs for every Claude model on Amazon Bedrock: schema-valid JSON, plus an EKS alert-triage example</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Wed, 28 Oct 2026 21:43:21 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/structured-outputs-for-every-claude-model-on-amazon-bedrock-schema-valid-json-plus-an-eks-2cpk</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/structured-outputs-for-every-claude-model-on-amazon-bedrock-schema-valid-json-plus-an-eks-2cpk</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published in Spanish on &lt;a href="https://blog.cloudacademy.ar/bedrock-structured-outputs-claude-json-schema/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;If you've put an LLM in the middle of an automated pipeline, you know the routine: ask for JSON in the prompt, add an example, beg for no extra text, and still get the occasional trailing comma, renamed field or "Sure, here's the JSON:" that breaks &lt;code&gt;json.loads&lt;/code&gt;. So you add validation, retries and a forgiving parser, and the glue code ends up bigger than the business logic.&lt;/p&gt;

&lt;p&gt;On October 20, 2026, AWS announced &lt;strong&gt;structured outputs for all Anthropic Claude models on Amazon Bedrock&lt;/strong&gt;, on both the &lt;code&gt;bedrock-runtime&lt;/code&gt; and &lt;code&gt;bedrock-mantle&lt;/code&gt; endpoints. You define a JSON Schema and the response follows it, because the constraint is applied during token generation rather than checked afterwards.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Two mechanisms: a &lt;strong&gt;JSON Schema output format&lt;/strong&gt; (shapes the model's response) and &lt;strong&gt;strict tool use&lt;/strong&gt; (&lt;code&gt;"strict": true&lt;/code&gt;, shapes tool-call arguments). You can use both in the same request.&lt;/li&gt;
&lt;li&gt;It's constrained decoding. The first request with a new schema compiles a grammar (up to ~90 seconds), which is cached for 24 hours per account.&lt;/li&gt;
&lt;li&gt;Every object needs &lt;code&gt;additionalProperties: false&lt;/code&gt;. Max 12 optional properties and 16 union-typed/&lt;code&gt;anyOf&lt;/code&gt; properties. No &lt;code&gt;minimum&lt;/code&gt;/&lt;code&gt;maximum&lt;/code&gt;, no &lt;code&gt;minLength&lt;/code&gt;/&lt;code&gt;maxLength&lt;/code&gt;, no recursive schemas.&lt;/li&gt;
&lt;li&gt;Incompatible with citations on Anthropic models (400 error).&lt;/li&gt;
&lt;li&gt;No separate charge announced: you pay model tokens, and output usually gets shorter.&lt;/li&gt;
&lt;li&gt;Still check &lt;code&gt;stop_reason&lt;/code&gt;: hitting &lt;code&gt;max_tokens&lt;/code&gt; truncates the JSON.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why this matters now
&lt;/h2&gt;

&lt;p&gt;Structured outputs on Bedrock isn't new. It went GA in February 2026 across several providers. The gap was coverage: the Claude 5 family wasn't supported for months. You could see it in the ecosystem. A July issue in Vercel's AI SDK reported Bedrock rejecting &lt;code&gt;strict&lt;/code&gt; for &lt;code&gt;claude-sonnet-5&lt;/code&gt; with a 400, so the SDK limited native structured outputs to Claude 4.x. The day of the announcement, a new issue asked to enable it for Claude 5. Lesson: Bedrock supporting it doesn't mean your SDK sends it.&lt;/p&gt;

&lt;p&gt;This launch closes that gap: every Claude model, both endpoints, plus batch inference, cross-Region inference and typed streaming events.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6opmt7f7o8jxxrw80z69.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F6opmt7f7o8jxxrw80z69.webp" alt="Structured outputs pipeline" width="800" height="295"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Bedrock validates the schema against a JSON Schema Draft 2020-12 subset (unsupported feature: immediate 400).&lt;/li&gt;
&lt;li&gt;A new schema compiles into a grammar, which can take up to ~90 seconds. AWS recommends a client read timeout of at least 120 seconds.&lt;/li&gt;
&lt;li&gt;The grammar is cached for 24 hours per account. Changing only &lt;code&gt;name&lt;/code&gt;/&lt;code&gt;description&lt;/code&gt; keeps the cache; structural changes don't.&lt;/li&gt;
&lt;li&gt;The model can only emit tokens that keep the output valid.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8555d1xteubvw5fw6lsa.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8555d1xteubvw5fw6lsa.webp" alt="Output format vs strict tool use" width="800" height="310"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Where the schema goes: &lt;code&gt;output_config.format&lt;/code&gt; for InvokeModel and the Messages API (a JSON object), and &lt;code&gt;outputConfig.textFormat&lt;/code&gt; for Converse (a JSON string inside &lt;code&gt;structure.jsonSchema.schema&lt;/code&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  Hands-on: alert triage on EKS with Claude Haiku 5.5
&lt;/h2&gt;

&lt;p&gt;The example: Alertmanager sends a webhook to a pod on EKS, the pod asks Claude Haiku 5.5 to classify the alert group against a schema, and the JSON decides who gets paged.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fl7r9kg92igaoc6jqiz35.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fl7r9kg92igaoc6jqiz35.webp" alt="Triage flow" width="798" height="212"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Haiku 5.5 is what AWS suggests for fast, repetitive work like classification. I'm using the global inference profile, &lt;code&gt;global.anthropic.claude-haiku-5-5&lt;/code&gt;, which is also the only way to reach Haiku 5.5 from sa-east-1.&lt;/p&gt;

&lt;p&gt;Permissions via EKS Pod Identity (trust &lt;code&gt;pods.eks.amazonaws.com&lt;/code&gt;, allow &lt;code&gt;bedrock:InvokeModel&lt;/code&gt; on the inference profile and foundation model ARNs), then:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws eks create-pod-identity-association &lt;span class="nt"&gt;--cluster-name&lt;/span&gt; production &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--namespace&lt;/span&gt; observability &lt;span class="nt"&gt;--service-account&lt;/span&gt; alert-triage &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--role-arn&lt;/span&gt; arn:aws:iam::123456789012:role/alert-triage-bedrock
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The schema keeps everything in &lt;code&gt;required&lt;/code&gt; to stay under the 12-optional limit, and uses &lt;code&gt;["string", "null"]&lt;/code&gt; where a value may be absent:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;TRIAGE_SCHEMA&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;object&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;properties&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;severity&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;string&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;enum&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;info&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;warning&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;critical&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]},&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;owner_team&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;string&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;enum&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;platform&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;payments&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;data&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;frontend&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;unknown&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]},&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;page_oncall&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;boolean&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;probable_cause&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;string&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;suggested_action&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;string&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;runbook_url&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;string&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;null&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;format&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;uri&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;related_alerts&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;array&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;items&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;string&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;}}&lt;/span&gt;
    &lt;span class="p"&gt;},&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;required&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;severity&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;owner_team&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;page_oncall&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;probable_cause&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                 &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;suggested_action&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;runbook_url&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;related_alerts&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;additionalProperties&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;False&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The call itself:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;bedrock&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;boto3&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;client&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;bedrock-runtime&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;region_name&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;sa-east-1&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                       &lt;span class="n"&gt;config&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nc"&gt;Config&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;read_timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;120&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;retries&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;max_attempts&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;mode&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;adaptive&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;}))&lt;/span&gt;

&lt;span class="n"&gt;resp&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;bedrock&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;invoke_model&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;modelId&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;global.anthropic.claude-haiku-5-5&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;dumps&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;anthropic_version&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;bedrock-2023-05-31&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;max_tokens&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;1024&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;system&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;SYSTEM&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;messages&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;role&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;user&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Alertmanager group:&lt;/span&gt;&lt;span class="se"&gt;\n&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;dumps&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;alerts&lt;/span&gt;&lt;span class="p"&gt;)}],&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;output_config&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;format&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;json_schema&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;schema&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;TRIAGE_SCHEMA&lt;/span&gt;&lt;span class="p"&gt;}},&lt;/span&gt;
    &lt;span class="p"&gt;}),&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;body&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;loads&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;resp&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;body&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="nf"&gt;read&lt;/span&gt;&lt;span class="p"&gt;())&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;stop_reason&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;end_turn&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="k"&gt;raise&lt;/span&gt; &lt;span class="nc"&gt;RuntimeError&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;incomplete response: &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;stop_reason&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;loads&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;next&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;text&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;b&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;text&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;There's no &lt;code&gt;jsonschema&lt;/code&gt; validation and no retry-on-malformed-JSON loop, because you don't need them. The one operational trick: &lt;strong&gt;warm up at pod start&lt;/strong&gt;. Make one dummy call before the readiness probe goes green (I use &lt;code&gt;periodSeconds: 5&lt;/code&gt; and &lt;code&gt;failureThreshold: 30&lt;/code&gt;), so grammar compilation happens during the rollout and not on the first real 3 a.m. alert. Run the receiver in parallel with your existing routing (&lt;code&gt;continue: true&lt;/code&gt;) for a few weeks and compare decisions.&lt;/p&gt;

&lt;p&gt;For agents that act, use strict tools. The &lt;code&gt;enum&lt;/code&gt; on &lt;code&gt;namespace&lt;/code&gt; and a &lt;code&gt;pattern&lt;/code&gt; on the deployment name constrain what the agent can ask for. Ranges like &lt;code&gt;1 &amp;lt;= replicas &amp;lt;= 20&lt;/code&gt; can't go in the schema, so they stay in code, and execution stays behind human approval.&lt;/p&gt;

&lt;h2&gt;
  
  
  Gotchas
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Short read timeouts kill the first request during grammar compilation.&lt;/li&gt;
&lt;li&gt;Low &lt;code&gt;max_tokens&lt;/code&gt; gives you truncated JSON even with structured outputs on.&lt;/li&gt;
&lt;li&gt;Going past 12 optional properties returns &lt;code&gt;Schema is too complex&lt;/code&gt; after up to ~90 seconds.&lt;/li&gt;
&lt;li&gt;A missing &lt;code&gt;additionalProperties: false&lt;/code&gt; in a nested object is an instant 400.&lt;/li&gt;
&lt;li&gt;Citations plus structured outputs on Claude is a 400.&lt;/li&gt;
&lt;li&gt;Every structural schema change pays compilation again. Version schemas like API contracts.&lt;/li&gt;
&lt;li&gt;Check what your SDK actually sends: some strip &lt;code&gt;strict&lt;/code&gt; for models they don't know about.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;No separate charge was announced; you pay the model's tokens. Anthropic's list prices per million tokens: Haiku 5.5 $0.10 input / $0.50 output (prompts up to 100K tokens), Sonnet 5.5 $2 / $10, Opus 5.5 $4 / $20. Check Bedrock's pricing table for your Region and profile. My estimate for 100,000 alert groups a month at ~1,500 input and 150 output tokens on Haiku: about &lt;strong&gt;$22.50/month&lt;/strong&gt;. Batch inference is 50% cheaper for non-urgent work and supports structured outputs. Prompt caching also helps, and Sonnet 5.5 cache reads on Bedrock cost half as much since October 7, 2026.&lt;/p&gt;

&lt;h2&gt;
  
  
  Region and data residency
&lt;/h2&gt;

&lt;p&gt;From São Paulo, Haiku 5.5 is only available through global cross-Region inference. That's fine for infrastructure alerts. For regulated data, check what a global profile implies first, and keep secrets and customer data out of alert annotations. Some geographic profiles don't support structured outputs for every model, so test with the exact profile your policy requires.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is it worth it?
&lt;/h2&gt;

&lt;p&gt;If you're using Claude on Bedrock for anything that ends in &lt;code&gt;json.loads&lt;/code&gt;: yes. Add &lt;code&gt;output_config&lt;/code&gt;, make the schema follow the rules, and delete your validation layer. For agents whose tools touch real systems, &lt;code&gt;strict: true&lt;/code&gt; should be the default from today.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;My take:&lt;/strong&gt; the feature has existed since February. What's new is that it no longer depends on the model. While some Claude models supported it and others didn't, every model switch risked breaking parsing, and libraries had to maintain allowlists. With full coverage, moving from Haiku to Sonnet goes back to being a quality and cost decision, not a compatibility one. For platform teams, the pattern generalizes: a small model, a strict schema, enums that map to real decisions, and a human approving anything with impact. It's the most boring and safest way to put AI into an ops workflow, and I mean that as a compliment.&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/10/claude-structured-outputs-bedrock/" rel="noopener noreferrer"&gt;What's New: structured outputs for Claude models&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/machine-learning/structured-outputs-for-anthropic-claude-models-on-amazon-bedrock/" rel="noopener noreferrer"&gt;AWS ML Blog: Structured outputs for Anthropic Claude models on Amazon Bedrock&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/machine-learning/structured-outputs-on-amazon-bedrock-schema-compliant-ai-responses" rel="noopener noreferrer"&gt;AWS ML Blog (Feb 2026): Structured outputs on Amazon Bedrock&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-structured-outputs.html" rel="noopener noreferrer"&gt;Docs: Claude structured outputs&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/bedrock/latest/userguide/structured-output.html" rel="noopener noreferrer"&gt;Docs: Get validated JSON results from models&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html#model-regions-anthropic" rel="noopener noreferrer"&gt;Anthropic model Region compatibility&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/vercel/ai/issues/23605" rel="noopener noreferrer"&gt;Vercel AI SDK issue #23605&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>bedrock</category>
      <category>ai</category>
      <category>kubernetes</category>
    </item>
    <item>
      <title>Amazon ECS Gateway: managed, shared ingress for ECS services (and the traps to avoid)</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Wed, 28 Oct 2026 21:42:30 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/amazon-ecs-gateway-managed-shared-ingress-for-ecs-services-and-the-traps-to-avoid-4gkj</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/amazon-ecs-gateway-managed-shared-ingress-for-ecs-services-and-the-traps-to-avoid-4gkj</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published in Spanish on &lt;a href="https://blog.cloudacademy.ar/amazon-ecs-gateway-ingress-administrado-alb-compartido/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;On October 20, 2026, AWS launched &lt;strong&gt;Amazon ECS Gateway&lt;/strong&gt;, an account-level resource that owns ingress for your ECS services. Instead of every service carrying its own Application Load Balancer, listener, target group, security group and certificate, a platform team defines a gateway once and services opt in with a single parameter on &lt;code&gt;create-service&lt;/code&gt; or &lt;code&gt;update-service&lt;/code&gt;. ECS provisions everything underneath and packs services onto shared ALBs, up to 25 services per load balancer.&lt;/p&gt;

&lt;p&gt;If you have ever written an internal Terraform module "so teams don't have to think about the ALB", you already know the problem this solves. The difference is that AWS now maintains that module, and it comes with compatibility rules, quotas and a couple of sharp edges worth knowing before you move production traffic.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;A gateway has its own ARN and provisions ALBs, HTTPS listeners, listener rules, target groups, security groups, ACM certificates and one hostname per service.&lt;/li&gt;
&lt;li&gt;Platform teams create gateways and associate them with clusters. App teams add a &lt;strong&gt;gateway connector&lt;/strong&gt; that names a port from the task definition.&lt;/li&gt;
&lt;li&gt;Each service gets &lt;code&gt;https://&amp;lt;unique-id&amp;gt;.ecs.&amp;lt;region&amp;gt;.on.aws&lt;/code&gt;. Up to 25 services share one ALB via host-based routing.&lt;/li&gt;
&lt;li&gt;HTTPS on port 443 only, &lt;code&gt;awsvpc&lt;/code&gt; network mode only (EC2, Fargate, ECS Managed Instances), same account only.&lt;/li&gt;
&lt;li&gt;No charge for the gateway itself. You pay for ALBs, public IPv4, logs, metrics and data transfer.&lt;/li&gt;
&lt;li&gt;Available in all commercial Regions where ECS runs.&lt;/li&gt;
&lt;li&gt;The biggest trap: &lt;code&gt;update-gateway&lt;/code&gt; replaces the entire configuration.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F47a8s7rgo22vvzeyfzfz.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F47a8s7rgo22vvzeyfzfz.webp" alt="Amazon ECS Gateway architecture" width="799" height="466"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;ECS Gateway architecture: platform declares ingress, each service gets a route on a shared ALB. Image: AWS.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  The model: gateway, association, connector, route
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Concept&lt;/th&gt;
&lt;th&gt;What it is&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Gateway&lt;/td&gt;
&lt;td&gt;Account-level resource, unique name per account and Region&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gateway-cluster association&lt;/td&gt;
&lt;td&gt;Many-to-many: up to 10,000 clusters per gateway, 5 gateways per cluster&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gateway connector&lt;/td&gt;
&lt;td&gt;Service parameter that connects it to every compatible gateway in its cluster&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gateway route&lt;/td&gt;
&lt;td&gt;One service-to-gateway connection, with status, URL and underlying resources&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Creating a gateway provisions nothing. Associating it with a cluster changes nobody's traffic. Only when a service with a connector deploys does ECS find a compatible gateway, place the service on an ALB with free capacity (or create one), and create the per-service certificate, hostname, rule, target group and security group.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbmqqpv3h3ie60myzk0ku.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbmqqpv3h3ie60myzk0ku.webp" alt="Responsibilities diagram" width="800" height="380"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;h3&gt;
  
  
  How ECS picks a gateway (read this twice)
&lt;/h3&gt;

&lt;p&gt;The connector &lt;strong&gt;does not select a gateway by name or ARN&lt;/strong&gt;. It attaches to every gateway in the cluster that:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;matches the connector's &lt;code&gt;accessTypes&lt;/code&gt; filter, if any (&lt;code&gt;PUBLIC&lt;/code&gt; = internet-facing, &lt;code&gt;PRIVATE&lt;/code&gt; = internal),&lt;/li&gt;
&lt;li&gt;is in the same VPC as the service, and&lt;/li&gt;
&lt;li&gt;covers every Availability Zone the service runs in.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;No match, and the operation fails. Several matches, and the service is connected to all of them. The docs flag the consequence explicitly: a service that must never be internet-reachable needs a &lt;code&gt;PRIVATE&lt;/code&gt; filter. Without it, if someone later associates an internet-facing gateway with the cluster, your internal service is exposed on its next deployment. Any redeploy will do.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdnzu9p8fosxucuq6b8s2.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdnzu9p8fosxucuq6b8s2.webp" alt="Console: create gateway" width="800" height="1191"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Creating a gateway in the ECS console. Image: AWS.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  What the platform team controls
&lt;/h2&gt;

&lt;p&gt;Immutable at creation: name, subnets (VPC and AZ coverage), IP address type (&lt;code&gt;ipv4&lt;/code&gt;, &lt;code&gt;dualstack&lt;/code&gt; default, &lt;code&gt;dualstack-without-public-ipv4&lt;/code&gt;) and scheme. The listener is always HTTPS:443.&lt;/p&gt;

&lt;p&gt;Mutable: &lt;code&gt;inboundSecurityGroup&lt;/code&gt;, &lt;code&gt;maxAlbCount&lt;/code&gt; (default 10, so 250 services), &lt;code&gt;webACLArn&lt;/code&gt;, &lt;code&gt;httpDesync&lt;/code&gt; (default &lt;code&gt;defensive&lt;/code&gt;), S3 access logs, &lt;code&gt;testTrafficRules&lt;/code&gt; (a header that routes requests to the new revision during blue/green or canary), &lt;code&gt;deletionProtection&lt;/code&gt; (default &lt;code&gt;true&lt;/code&gt;) and &lt;code&gt;propagateTags&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Raise &lt;code&gt;maxAlbCount&lt;/code&gt; before you need it: a service that opts in at capacity fails its deployment, and raising the limit afterwards does not retry it.&lt;/p&gt;
&lt;h2&gt;
  
  
  Hands-on with the AWS CLI
&lt;/h2&gt;

&lt;p&gt;Infrastructure role:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws iam create-role &lt;span class="nt"&gt;--role-name&lt;/span&gt; ecsInfrastructureRoleForGateways &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--assume-role-policy-document&lt;/span&gt; &lt;span class="s1"&gt;'{"Version":"2012-10-17","Statement":[{"Effect":"Allow","Principal":{"Service":"ecs.amazonaws.com"},"Action":"sts:AssumeRole"}]}'&lt;/span&gt;
aws iam attach-role-policy &lt;span class="nt"&gt;--role-name&lt;/span&gt; ecsInfrastructureRoleForGateways &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--policy-arn&lt;/span&gt; arn:aws:iam::aws:policy/AmazonECSInfrastructureRoleForGatewayHostSharedALB
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Gateway with access logs and a test-traffic header:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws ecs create-gateway &lt;span class="nt"&gt;--name&lt;/span&gt; platform-public &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--infrastructure-role-arn&lt;/span&gt; arn:aws:iam::123456789012:role/ecsInfrastructureRoleForGateways &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--configuration&lt;/span&gt; &lt;span class="s1"&gt;'{"hostSharedAlb":{"scheme":"internet-facing",
    "networkConfiguration":{"subnets":["subnet-aaa111","subnet-bbb222","subnet-ccc333"],"ipAddressType":"dualstack"},
    "maxAlbCount":20,"httpDesync":"monitor",
    "accessLogs":{"destination":{"s3":{"bucket":"my-alb-logs","prefix":"gateway"}}},
    "testTrafficRules":{"header":{"name":"x-test-traffic","values":["canary"]}}}}'&lt;/span&gt;

aws ecs associate-gateway-with-cluster &lt;span class="nt"&gt;--gateway&lt;/span&gt; platform-public &lt;span class="nt"&gt;--cluster&lt;/span&gt; production
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Name the port in the task definition (&lt;code&gt;portMappings[].name&lt;/code&gt;, e.g. &lt;code&gt;web&lt;/code&gt;), then connect the service:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws ecs create-service &lt;span class="nt"&gt;--cluster&lt;/span&gt; production &lt;span class="nt"&gt;--service-name&lt;/span&gt; payments-api &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--task-definition&lt;/span&gt; payments-api:7 &lt;span class="nt"&gt;--desired-count&lt;/span&gt; 3 &lt;span class="nt"&gt;--launch-type&lt;/span&gt; FARGATE &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--network-configuration&lt;/span&gt; &lt;span class="s1"&gt;'awsvpcConfiguration={subnets=[subnet-priv1,subnet-priv2,subnet-priv3],securityGroups=[sg-0123abcd],assignPublicIp=DISABLED}'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--deployment-configuration&lt;/span&gt; &lt;span class="s1"&gt;'{"strategy": "BLUE_GREEN"}'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--gateway-connectors&lt;/span&gt; &lt;span class="s1"&gt;'[{"portName":"web","filters":{"accessTypes":["PUBLIC"]},"healthCheck":{"healthCheckPath":"/health"}}]'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Get the URL from the service revision:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;DEPLOY_ARN&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;aws ecs list-service-deployments &lt;span class="nt"&gt;--cluster&lt;/span&gt; production &lt;span class="nt"&gt;--service&lt;/span&gt; payments-api &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--query&lt;/span&gt; &lt;span class="s1"&gt;'serviceDeployments[0].serviceDeploymentArn'&lt;/span&gt; &lt;span class="nt"&gt;--output&lt;/span&gt; text&lt;span class="si"&gt;)&lt;/span&gt;
&lt;span class="nv"&gt;REV_ARN&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;aws ecs describe-service-deployments &lt;span class="nt"&gt;--service-deployment-arns&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$DEPLOY_ARN&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--query&lt;/span&gt; &lt;span class="s1"&gt;'serviceDeployments[0].targetServiceRevision.arn'&lt;/span&gt; &lt;span class="nt"&gt;--output&lt;/span&gt; text&lt;span class="si"&gt;)&lt;/span&gt;
aws ecs describe-service-revisions &lt;span class="nt"&gt;--service-revision-arns&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$REV_ARN&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--query&lt;/span&gt; &lt;span class="s1"&gt;'serviceRevisions[0].ecsManagedResources.gatewayConnections'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Wait for &lt;code&gt;ACTIVE&lt;/code&gt;. &lt;code&gt;CREATE_FAILED&lt;/code&gt; won't fix itself by waiting.&lt;/p&gt;

&lt;h2&gt;
  
  
  Gotchas
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;update-gateway&lt;/code&gt; replaces everything.&lt;/strong&gt; Send only &lt;code&gt;maxAlbCount&lt;/code&gt; and your WAF is detached, logs are turned off, the test header disappears and deletion protection resets. Always &lt;code&gt;describe-gateway&lt;/code&gt; first.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Health check changes replace the target group&lt;/strong&gt;, which fails on &lt;code&gt;ROLLING&lt;/code&gt; deployments. Get it right on first connect, or use &lt;code&gt;BLUE_GREEN&lt;/code&gt;/&lt;code&gt;CANARY&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Default subnets&lt;/strong&gt; come from the default VPC, which rarely matches your services.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No self-repair&lt;/strong&gt; if someone edits managed resources by hand, and &lt;strong&gt;no consolidation&lt;/strong&gt; of half-empty ALBs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not for Express Mode services&lt;/strong&gt;, one connector per service, same account only.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Terraform&lt;/strong&gt;: support is still an open request (issue #50573, filed October 21).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Custom domains&lt;/strong&gt; aren't covered in the current docs; services get &lt;code&gt;on.aws&lt;/code&gt; hostnames.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;No gateway charge. Reference prices in us-east-1: ALB at $0.0225/hour plus $0.008 per LCU-hour; public IPv4 at $0.005/hour; public ACM certificates are free.&lt;/p&gt;

&lt;p&gt;My own back-of-the-envelope for fixed costs only (730 hours, three AZs): 30 services with one ALB each come to about &lt;strong&gt;$821/month&lt;/strong&gt; in ALB hours plus public IPv4. The same 30 services on a gateway fit in two ALBs, about &lt;strong&gt;$55/month&lt;/strong&gt;. LCU charges follow traffic and don't disappear. And with 25 rules on a shared listener you go past the 10 free rules, so the rule-evaluation dimension starts counting. If clients support IPv6, &lt;code&gt;dualstack-without-public-ipv4&lt;/code&gt; removes the IPv4 line entirely.&lt;/p&gt;

&lt;h2&gt;
  
  
  Availability and what to check
&lt;/h2&gt;

&lt;p&gt;It's in every commercial Region with ECS, including &lt;strong&gt;sa-east-1 (São Paulo)&lt;/strong&gt;. Everything it creates lives in your account and Region. Multi-account setups need one gateway per account. Lock down &lt;code&gt;associate-gateway-with-cluster&lt;/code&gt; with IAM, because that action decides what becomes exposed on other teams' next deployments.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is it worth it?
&lt;/h2&gt;

&lt;p&gt;For teams running dozens of HTTP services on ECS and maintaining shared ALBs by hand or through internal modules: yes. The savings are real, and the split between platform and app teams is the one platform teams want. If you rely on path-based rules, custom domains, mTLS, NLBs or ALB-native OIDC authentication, wait: none of that shows up in the current docs.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;My take:&lt;/strong&gt; the shared ALB isn't the news; that was always possible. The news is that AWS now owns the full lifecycle: create, place, scale out to more ALBs, clean up. It mirrors what the AWS Load Balancer Controller's IngressGroups and the Kubernetes Gateway API do on EKS, with separate roles for whoever defines infrastructure and whoever publishes routes. What worries me is that "connect to every compatible gateway" plus "associations apply on next deploy" turns a platform action into an exposure change for someone else's service. My rule: every connector gets an explicit &lt;code&gt;accessTypes&lt;/code&gt; filter, always, and very few people can associate gateways.&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/10/announcing-amazon-ecs-gateway/" rel="noopener noreferrer"&gt;What's New announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/aws/introducing-amazon-ecs-gateway-a-fully-managed-ingress-solution-for-ecs-services/" rel="noopener noreferrer"&gt;AWS News Blog: Introducing Amazon ECS Gateway&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/AmazonECS/latest/developerguide/gateway-overview.html" rel="noopener noreferrer"&gt;Docs: ECS Gateway overview&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/AmazonECS/latest/developerguide/gateway-managing.html" rel="noopener noreferrer"&gt;Docs: create and manage a gateway&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/AmazonECS/latest/developerguide/gateway-connect-service.html" rel="noopener noreferrer"&gt;Docs: connect services to a gateway&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/elasticloadbalancing/pricing/" rel="noopener noreferrer"&gt;Elastic Load Balancing pricing&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/hashicorp/terraform-provider-aws/issues/50573" rel="noopener noreferrer"&gt;Terraform AWS provider issue #50573&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>ecs</category>
      <category>containers</category>
      <category>devops</category>
    </item>
    <item>
      <title>Strands Decider 2B: AWS's open source model that doesn't generate text, it decides</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Tue, 13 Oct 2026 21:25:40 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/strands-decider-2b-awss-open-source-model-that-doesnt-generate-text-it-decides-48gm</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/strands-decider-2b-awss-open-source-model-that-doesnt-generate-text-it-decides-48gm</guid>
      <description>&lt;p&gt;A lot of the LLM calls inside an agent aren't about generating text. They're tiny decisions: which team owns this ticket? Did the user actually say that tool argument, or did the model make it up? Is this message urgent? Cheap model or expensive model? For each one, we ship the full context to a frontier model, wait a second or two, and parse free text that comes with no measure of how confident the model is.&lt;/p&gt;

&lt;p&gt;Strands Labs, the experimental arm of the Strands Agents team at AWS, released &lt;strong&gt;Strands Decider 2B&lt;/strong&gt;, an Apache 2.0 model built for exactly that gap. It doesn't generate text. It takes a state plus a set of questions and, for each one, picks from closed options and returns a calibrated probability. It runs locally, answers in about 115 ms median on a desktop GPU, and plugs into Strands Agents' intervention system to review tool calls before they execute.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;A &lt;strong&gt;decision model&lt;/strong&gt; (the authors call it a "system one" model): picks options, answers yes/no, or scores on a scale, always with a probability. No text generation.&lt;/li&gt;
&lt;li&gt;Architecture: the &lt;strong&gt;Qwen3.5-2B-Base&lt;/strong&gt; torso with its LM head removed, a ~1M-parameter &lt;strong&gt;pointer head&lt;/strong&gt;, and rank-16 LoRA adapters. One forward pass, no decoding loop.&lt;/li&gt;
&lt;li&gt;First release latency: &lt;strong&gt;115 ms median&lt;/strong&gt; (299 ms p95) on an RTX 3090; 153 ms on an M3 Pro.&lt;/li&gt;
&lt;li&gt;On JevBench (231 public tasks) the current reference model gets 180 right (0.779). Gemma 4 variants reach 207 with the 26B model.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;pip install strands-decider&lt;/code&gt;, then use it from the CLI or as a local HTTP server (&lt;code&gt;/v1/systemone&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;It's an experiment, not a managed AWS service. You run it.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why now
&lt;/h2&gt;

&lt;p&gt;The "decision model" idea got attention in September when TypeSafe AI came out of stealth with &lt;strong&gt;Jev&lt;/strong&gt;, a proprietary model that returns typed decisions instead of text, available via waitlist at $0.042 per million input tokens. Strands credits Jev for putting the approach on the map and evaluates against JevBench. The difference is distribution: Jev is closed; Strands Decider ships weights, code and the full training recipe.&lt;/p&gt;

&lt;p&gt;There's also a design argument. Ask a chat model "are your own tool arguments grounded?" and it can be talked into yes. A classifier that returns a number your &lt;code&gt;if&lt;/code&gt; statement branches on doesn't negotiate.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fowf15x5x0nvifi04thby.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fowf15x5x0nvifi04thby.webp" alt="The model classifies, your code decides" width="800" height="246"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;The decider keeps the transformer torso of Qwen3.5-2B-Base, throws away the vocabulary head, and replaces it with a pointer head. The input has a fixed format: state, question, numbered options, and an &lt;code&gt;&amp;lt;answer&amp;gt;&lt;/code&gt; marker. The head scores each option's final hidden state against the hidden state at &lt;code&gt;&amp;lt;answer&amp;gt;&lt;/code&gt;, a softmax turns those logits into probabilities, done. No tokens generated.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm2acxdovtrtc2jdev2wo.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm2acxdovtrtc2jdev2wo.webp" alt="Strands Decider 2B architecture. Image: Strands Agents blog (AWS)" width="800" height="502"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Two practical consequences: the head has no per-option parameters, so it can't learn "option 1 is usually right" and you can rename options without retraining; and only the LoRA adapters and the head are trained, which is why the repo can publish the recipe (about 11 hours on one RTX 3090, or ~70 minutes on eight H100s).&lt;/p&gt;

&lt;p&gt;Three question types are supported: &lt;code&gt;noul&lt;/code&gt; (yes/no, returns P(yes)), &lt;code&gt;choice&lt;/code&gt; (one of N), and &lt;code&gt;score&lt;/code&gt; (a level on an ordered rubric). You can ask many questions about one state in a single request; the state is encoded once.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;JevBench (of 231)&lt;/th&gt;
&lt;th&gt;Brier (lower is better)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;2B hobson-v19 (launch)&lt;/td&gt;
&lt;td&gt;167&lt;/td&gt;
&lt;td&gt;0.342&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2B hobson-v21&lt;/td&gt;
&lt;td&gt;176&lt;/td&gt;
&lt;td&gt;0.323&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2B qwen3.5-v1-2610 (reference)&lt;/td&gt;
&lt;td&gt;180&lt;/td&gt;
&lt;td&gt;0.280&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;E4B gemma4-v1-2610&lt;/td&gt;
&lt;td&gt;187&lt;/td&gt;
&lt;td&gt;0.243&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;26B-A4B gemma4-v1-2610&lt;/td&gt;
&lt;td&gt;207&lt;/td&gt;
&lt;td&gt;0.162&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The README itself warns that gaps under ~10 tasks between single runs aren't reliable. For v19, Techstrong.ai reports 100% on easy tasks and 50.5% on hard ones: great at routine, mediocre at complex, which is the pitch.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fo6yoks8llc15o5rgl4zb.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fo6yoks8llc15o5rgl4zb.webp" alt="Accuracy vs. Brier score across training versions. Image: Strands Agents blog (AWS)" width="800" height="654"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Faoyx4tad528pexins6t6.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Faoyx4tad528pexins6t6.webp" alt="Latency vs. prompt length on an RTX 3090. Image: Strands Agents blog (AWS)" width="800" height="462"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;strands-decider

strands-decider ask StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--state&lt;/span&gt; &lt;span class="s2"&gt;"Help! My payouts have been failing for 3 days! "&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--choice&lt;/span&gt; &lt;span class="s2"&gt;"Which team should handle this?=billing,sales,retail"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--noul&lt;/span&gt; &lt;span class="s2"&gt;"Does this convey urgency?"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--score&lt;/span&gt; &lt;span class="s2"&gt;"How frustrated is the writer?=calm,frustrated,depressed"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Sample output from the README (CPU run): &lt;code&gt;billing&lt;/code&gt; at 0.773, urgency 0.858, and a frustration score of 1.03 with confidence 0.509. That last number is the model telling you it isn't sure, which is information an LLM doesn't give you.&lt;/p&gt;

&lt;p&gt;As a server:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;strands-decider serve StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 &lt;span class="nt"&gt;--port&lt;/span&gt; 8000
curl &lt;span class="nt"&gt;-s&lt;/span&gt; localhost:8000/v1/systemone &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s1"&gt;'content-type: application/json'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{"state": "Help! My payouts have been failing for 3 days!",
       "questions": {"is_urgent": {"type": "noul", "instructions": "Does this convey urgency?"}}}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The server binds to &lt;code&gt;127.0.0.1&lt;/code&gt; and &lt;strong&gt;has no authentication&lt;/strong&gt;. Don't put it behind a public Service.&lt;/p&gt;

&lt;h3&gt;
  
  
  Gating tool calls in a Strands agent
&lt;/h3&gt;

&lt;p&gt;This is a trimmed version of &lt;code&gt;examples/strands/&lt;/code&gt; from the repo. The agent has an eager system prompt that makes it guess a city; the decider catches it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;strands&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;tool&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;strands.interventions&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Guide&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;InterventionHandler&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;Proceed&lt;/span&gt;

&lt;span class="n"&gt;YES&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;0.45&lt;/span&gt;  &lt;span class="c1"&gt;# illustrative threshold; tune it on your own traffic
&lt;/span&gt;
&lt;span class="k"&gt;class&lt;/span&gt; &lt;span class="nc"&gt;ToolCallReviewer&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;InterventionHandler&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;strands-decider-tool-call-reviewer&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;

    &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;__init__&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;decider&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;_decider&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;decider&lt;/span&gt;

    &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;before_tool_call&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;kwargs&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="n"&gt;state&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;build_state&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;tool_use&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;  &lt;span class="c1"&gt;# conversation + proposed call
&lt;/span&gt;        &lt;span class="n"&gt;answers&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;_decider&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;ask&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;state&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;QUESTIONS&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;  &lt;span class="c1"&gt;# args_grounded, premature (noul)
&lt;/span&gt;        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;answers&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;args_grounded&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;noul&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="n"&gt;YES&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nc"&gt;Guide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feedback&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;The arguments look guessed. Ask the user to confirm them.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;answers&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;premature&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;noul&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="n"&gt;YES&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nc"&gt;Guide&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;feedback&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Too early to call this tool. Clarify with the user first.&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nc"&gt;Proceed&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;Guide&lt;/code&gt; doesn't block the call; it hands the turn back to the model with feedback, and the agent ends up asking "which city?" instead of reporting Seattle's weather. Strands interventions also support &lt;code&gt;Deny&lt;/code&gt; and &lt;code&gt;Confirm&lt;/code&gt; (human approval).&lt;/p&gt;

&lt;h3&gt;
  
  
  Running it on Amazon EKS
&lt;/h3&gt;

&lt;p&gt;Because the server only listens on localhost, the natural Kubernetes pattern is a &lt;strong&gt;sidecar in the agent's pod&lt;/strong&gt;. A native sidecar (an init container with &lt;code&gt;restartPolicy: Always&lt;/code&gt;) starts before the agent and lives as long as the pod. Two gotchas: there's no published container image, so build your own and bake the weights in; and &lt;code&gt;httpGet&lt;/code&gt; probes will always fail because the kubelet hits the pod IP, not 127.0.0.1, so use an &lt;code&gt;exec&lt;/code&gt; probe against &lt;code&gt;http://127.0.0.1:&amp;lt;port&amp;gt;/health&lt;/code&gt;. A g6-family node (NVIDIA L4, 24 GB) is plenty for the 2B model.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fk4stz6vv7zczb8ojvsdb.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fk4stz6vv7zczb8ojvsdb.webp" alt="Decider as a sidecar on Amazon EKS" width="800" height="284"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Cost
&lt;/h2&gt;

&lt;p&gt;The model is free; you pay for hardware. A g6.xlarge is about $0.805/hour on demand in us-east-1, roughly $588/month always on. Back-of-the-envelope (mine): ~8 sequential decisions/second at ~115 ms is on the order of 20M decisions a month; at 500 tokens each, that volume on a per-token service like Jev would be around $450. Self-hosting wins with sustained volume or privacy requirements; at low volume an idle GPU costs more. Start on CPU, share the GPU, and use the decider as a first-pass filter that escalates low-confidence cases to the LLM.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations worth knowing
&lt;/h2&gt;

&lt;p&gt;Per the model card: rephrased questions often get the same answer, long multi-step documents are the weak spot, &lt;code&gt;score&lt;/code&gt; and &lt;code&gt;noul&lt;/code&gt; transfer poorly to unfamiliar rubrics, and calibration was fitted on short classification, so validate thresholds on your own traffic. Training data and examples are English.&lt;/p&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;The benchmark isn't the point (231 tasks is noisy). The point is the separation: the model classifies and returns a number, your code decides. That turns decisions hidden inside prompts into something you can test, version and audit with explicit thresholds. It's the same logic platform teams apply with Kubernetes admission policies: a deterministic checkpoint before something runs. It's not production-ready, and the official example says its thresholds are illustrative. But I'd run it as a tool-call guardrail for an internal agent, in a pod with a shared GPU, and compare its decisions against what your LLM decides today.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published in Spanish on &lt;a href="https://blog.cloudacademy.ar/strands-decider-2b-modelo-de-decision-open-source-para-agentes/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;, with the full Kubernetes manifest and the complete intervention example.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://strandsagents.com/blog/introducing-strands-decider/" rel="noopener noreferrer"&gt;Introducing Strands Decider 2B&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/strands-labs/strands-decider" rel="noopener noreferrer"&gt;strands-labs/strands-decider on GitHub&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/strands-labs/strands-decider/tree/main/examples/strands" rel="noopener noreferrer"&gt;Strands agent intervention example&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/StrandsAgents/strands-decider-2B-hobson-v19" rel="noopener noreferrer"&gt;Model card on Hugging Face&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://techstrong.ai/articles/aws-explores-decision-models-with-strands-decider-2b/" rel="noopener noreferrer"&gt;Techstrong.ai coverage&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.vktr.com/ai-platforms/chatgpt-cocreator-launches-typesafe-ai-with-jev/" rel="noopener noreferrer"&gt;VKTR: TypeSafe AI launches Jev&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>ai</category>
      <category>agents</category>
      <category>kubernetes</category>
    </item>
    <item>
      <title>AWS Lambda Web Functions: your HTTP server on Lambda, with real concurrency and active-CPU billing</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Tue, 13 Oct 2026 21:18:40 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/aws-lambda-web-functions-your-http-server-on-lambda-with-real-concurrency-and-active-cpu-billing-36eb</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/aws-lambda-web-functions-your-http-server-on-lambda-with-real-concurrency-and-active-cpu-billing-36eb</guid>
      <description>&lt;p&gt;On October 12, AWS launched &lt;strong&gt;Lambda Web Functions&lt;/strong&gt;, and it's the biggest change to Lambda's programming model in years. Instead of writing a handler that receives a JSON event, you ship a plain HTTP server (Express, Hono, Fastify, React Router) that listens on a port. Lambda puts a managed HTTPS endpoint in front of it, scales it with traffic (down to zero), and bills you for the CPU time you actually burn, not for the time your request spent waiting on Postgres.&lt;/p&gt;

&lt;p&gt;Along with the launch, the existing functions were renamed &lt;strong&gt;Lambda Event Functions&lt;/strong&gt; (naming only, no code changes), so Lambda now comes in three shapes: Event Functions, Web Functions, and MicroVMs.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffqlywez4a3hvcxq1xkbi.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffqlywez4a3hvcxq1xkbi.webp" alt="The three shapes of AWS Lambda" width="800" height="338"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Your server listens on the &lt;code&gt;host:port&lt;/code&gt; Lambda passes in &lt;code&gt;AWS_LAMBDA_HTTP_ENDPOINT&lt;/code&gt;. No adapter, no wrapper.&lt;/li&gt;
&lt;li&gt;Each execution environment serves &lt;strong&gt;many requests at once&lt;/strong&gt;: up to 128, default 64.&lt;/li&gt;
&lt;li&gt;Billing is &lt;strong&gt;active CPU + memory consumed while the environment lives + requests&lt;/strong&gt;. In us-east-1 (Arm): $0.0000276944 per vCPU-second, $0.0000036667 per GB-second, $0.30 per million requests.&lt;/li&gt;
&lt;li&gt;One-command deploy: &lt;code&gt;aws lambda-web deploy --name my-app --code ./my-app&lt;/code&gt;. Add &lt;code&gt;--regions&lt;/code&gt; for a global endpoint with latency-based routing.&lt;/li&gt;
&lt;li&gt;Immutable revisions plus weighted endpoints give you canary, blue/green and rollback out of the box.&lt;/li&gt;
&lt;li&gt;Node.js 24 on Arm64 only, 2 GB per environment, 3–900 s request timeout.&lt;/li&gt;
&lt;li&gt;Available in us-east-1, us-east-2, us-west-2, ap-northeast-1 and eu-west-1. Not in São Paulo yet.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why this matters
&lt;/h2&gt;

&lt;p&gt;Running web apps on Lambda was always possible, but awkward. You wrapped your framework with &lt;code&gt;serverless-http&lt;/code&gt; or the Lambda Web Adapter, and each environment still handled exactly one request. If your API took 200 ms because it was waiting on a database, you paid for 200 ms of compute while the CPU idled most of that time. A burst of 50 concurrent requests meant up to 50 environments and 50 cold starts.&lt;/p&gt;

&lt;p&gt;That model fits short, event-driven functions. It fits LLM-backed APIs terribly: a request that streams tokens for three seconds while the CPU mostly waits is the worst case for wall-clock billing and single-request isolation. Web Functions is AWS's answer, and it's no coincidence the launch post lists streaming inference endpoints among the first use cases.&lt;/p&gt;

&lt;h2&gt;
  
  
  The resource model: functions, revisions, endpoints
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Web Function&lt;/strong&gt;: the top-level resource.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Revision&lt;/strong&gt;: an immutable snapshot of code and configuration (timeout, per-environment concurrency, env vars, execution role). Any change means a new revision, and a new revision doesn't inherit settings from the previous one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Endpoint&lt;/strong&gt;: the HTTPS entry point, with its own domain, auth type and scaling limits. It routes to one or two revisions by weight.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flwg3awrncd0djprmr9m7.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flwg3awrncd0djprmr9m7.webp" alt="Web Function resource model: prod endpoint split 90/10, staging endpoint at 100% on revision 2. Image: AWS documentation" width="800" height="286"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Endpoints come in three types, fixed at creation: &lt;code&gt;HomeRegion&lt;/code&gt; (one Region, one domain), &lt;code&gt;MultiRegion&lt;/code&gt; (several Regions behind a global &lt;code&gt;*.lambda-web.global.on.aws&lt;/code&gt; domain with best-effort latency routing), and &lt;code&gt;PerRegion&lt;/code&gt; (one domain per Region, you route).&lt;/p&gt;

&lt;p&gt;Each environment goes through init, ready and shutdown. Your server must start listening &lt;strong&gt;within 10 seconds&lt;/strong&gt;; listening on the port is the readiness signal, there's no separate health check. On shutdown you get &lt;code&gt;SIGTERM&lt;/code&gt; and &lt;strong&gt;one second&lt;/strong&gt; to clean up.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Feqkvpaivqgs8espa99x3.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Feqkvpaivqgs8espa99x3.webp" alt="Execution environment lifecycle. Image: AWS documentation" width="799" height="251"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Concurrency and scaling
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Setting&lt;/th&gt;
&lt;th&gt;Where&lt;/th&gt;
&lt;th&gt;Range&lt;/th&gt;
&lt;th&gt;Default&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;maxConcurrencyPerEnvironment&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Revision&lt;/td&gt;
&lt;td&gt;1–128&lt;/td&gt;
&lt;td&gt;64&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;maxEnvironments&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Endpoint, per Region&lt;/td&gt;
&lt;td&gt;2–10,000 (optional)&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;rateLimit&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Endpoint, per Region&lt;/td&gt;
&lt;td&gt;0–10,000 RPS (optional)&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;timeoutSeconds&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Revision&lt;/td&gt;
&lt;td&gt;3–900&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Account quotas apply per Region even if you set nothing: 10,000 RPS per endpoint, 100,000 RPS per account, and 2,000 vCPUs. Each environment uses 2 vCPUs, so that's at most 1,000 environments per Region. Lambda keeps headroom to absorb gradual increases; bursts larger than the headroom wait on cold starts. When you hit a limit you get a 429 (not queued); a &lt;code&gt;rateLimit&lt;/code&gt; of 0 returns &lt;code&gt;503 Platform.EndpointDisabled&lt;/code&gt;, which doubles as a kill switch.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flgzr1tqpzp35tlsn2x2n.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flgzr1tqpzp35tlsn2x2n.webp" alt="Scaling headroom vs. traffic burst. Image: AWS documentation" width="799" height="391"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;

&lt;p&gt;The only code change versus running locally is where you listen:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;express&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;require&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;express&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;app&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;express&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;span class="nx"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;/&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;req&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;res&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="nx"&gt;res&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;json&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt; &lt;span class="na"&gt;message&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;Hello from Lambda Web Functions&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt; &lt;span class="p"&gt;}));&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nx"&gt;host&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;port&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;env&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;AWS_LAMBDA_HTTP_ENDPOINT&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;localhost:3000&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;split&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;:&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="nx"&gt;app&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;listen&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;port&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;host&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="nx"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;on&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;SIGTERM&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="nx"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;exit&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;));&lt;/span&gt; &lt;span class="c1"&gt;// you get ~1s&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Deploy (make sure your AWS CLI v2 has &lt;code&gt;aws lambda-web&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws lambda-web deploy &lt;span class="nt"&gt;--name&lt;/span&gt; my-api &lt;span class="nt"&gt;--code&lt;/span&gt; ./my-api &lt;span class="nt"&gt;--auth-type&lt;/span&gt; IamAuth
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Note that &lt;strong&gt;&lt;code&gt;deploy&lt;/code&gt; creates an &lt;code&gt;ApplicationManaged&lt;/code&gt; (public, no auth) endpoint by default&lt;/strong&gt;. Great for hello world, not for production accounts. The docs include an SCP that denies creating such endpoints via the &lt;code&gt;lambda:Request/WebFunctionAuthType&lt;/code&gt; condition key.&lt;/p&gt;

&lt;p&gt;Canary and rollback are just weight updates:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws lambda-web update-web-function-endpoint &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--function-name&lt;/span&gt; my-api &lt;span class="nt"&gt;--endpoint-name&lt;/span&gt; prod &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--revision-weights&lt;/span&gt; &lt;span class="s1"&gt;'[{"revisionId": "rev-abc12", "weight": 90}, {"revisionId": "rev-def45", "weight": 10}]'&lt;/span&gt;

&lt;span class="c"&gt;# rollback (include the mode so it works from any state)&lt;/span&gt;
aws lambda-web update-web-function-endpoint &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--function-name&lt;/span&gt; my-api &lt;span class="nt"&gt;--endpoint-name&lt;/span&gt; prod &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--auto-deployment-mode&lt;/span&gt; Disabled &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--revision-weights&lt;/span&gt; &lt;span class="s1"&gt;'[{"revisionId": "rev-abc12", "weight": 100}]'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm7nx9u3mwtyymz963f9v.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm7nx9u3mwtyymz963f9v.webp" alt="Canary deployment with revision weights. Image: AWS documentation" width="800" height="286"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Protect your downstream database without a new revision:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws lambda-web update-web-function-endpoint &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--function-name&lt;/span&gt; my-api &lt;span class="nt"&gt;--endpoint-name&lt;/span&gt; prod &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--scaling-config&lt;/span&gt; &lt;span class="s1"&gt;'{"maxEnvironments": 50}'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--throttle-config&lt;/span&gt; &lt;span class="s1"&gt;'{"rateLimit": 8000}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For a custom domain and AWS WAF, put CloudFront in front with &lt;strong&gt;Origin Access Control&lt;/strong&gt; (origin type &lt;code&gt;lambda&lt;/code&gt;, signing behavior &lt;code&gt;always-amz-auth&lt;/code&gt;) against an &lt;code&gt;IamAuth&lt;/code&gt; endpoint. CloudFront origin support launched the same day at no extra cost.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdaslzen7g23uhzui7abf.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdaslzen7g23uhzui7abf.webp" alt="CloudFront + WAF in front of an IamAuth Web Function endpoint. Image: AWS documentation" width="799" height="251"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Outbound VPC access uses a network connector (&lt;code&gt;aws lambda-core create-network-connector&lt;/code&gt;) referenced by version-qualified ARN from the revision. It only works with &lt;code&gt;HomeRegion&lt;/code&gt; endpoints, and you still need a NAT gateway for internet egress.&lt;/p&gt;

&lt;h3&gt;
  
  
  Gotchas
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;"Active" doesn't mean "working."&lt;/strong&gt; Lambda doesn't check that your server answers. Always smoke-test after deploying.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hardcoded ports&lt;/strong&gt; never become ready. Read &lt;code&gt;AWS_LAMBDA_HTTP_ENDPOINT&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Global mutable state&lt;/strong&gt; is now a concurrency bug, just like in a container.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;2 GB is shared&lt;/strong&gt; by heap, buffers and every in-flight request. V8's default heap is 1,024 MB.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;DNS is cached&lt;/strong&gt; per environment; ship a new revision after private DNS changes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Signed requests&lt;/strong&gt; to &lt;code&gt;IamAuth&lt;/code&gt; endpoints must include &lt;code&gt;X-Amz-Content-SHA256&lt;/code&gt;, even with no body.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;delete-web-function&lt;/code&gt; leaves the default IAM role, S3 packages and log groups behind.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Pricing, honestly
&lt;/h2&gt;

&lt;p&gt;AWS's own examples assume one environment for a 30-day month: an SSR app with 15M requests costs $25.78; an I/O-heavy API with 10M requests (200 ms, 50 ms CPU) costs $19.61; an LLM streaming endpoint with 5M requests of 3 s (200 ms CPU) costs $33.94.&lt;/p&gt;

&lt;p&gt;I ran the same workloads through Event Function pricing (x86 rate listed on the same page, my assumption of 512 MB, no free tier): roughly $18.67 for the I/O API, $21.75 for SSR, and &lt;strong&gt;about $126 for LLM streaming&lt;/strong&gt;. So Web Functions isn't automatically cheaper for short requests (the vCPU-second rate is high and requests cost 50% more), but for long, I/O-bound requests like token streaming it's a different order of magnitude.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh77buaxo25ozrgkdjey4.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh77buaxo25ozrgkdjey4.webp" alt="What gets billed in a 200 ms request" width="800" height="300"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What to check before adopting
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Regions:&lt;/strong&gt; five at launch, no South America. For users in Argentina or Brazil, us-east-1 is the closest option.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Data residency:&lt;/strong&gt; multi-Region endpoints replicate code, and request metrics and logs are centralized asynchronously in the home Region.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cross-Region data transfer:&lt;/strong&gt; a global endpoint serving from Tokyo but querying a database in Virginia pays in latency and transfer.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Maturity:&lt;/strong&gt; one runtime, brand-new APIs, IaC tooling still catching up. LocalStack 2026.10.0 already emulates it for local testing.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;The interesting part isn't "Express on Lambda"; we could already do that. It's that AWS changed both the billing unit and the isolation unit of its most used serverless service. For ten years Lambda made us think in one request per environment and pay for waiting time. That made sense for short event handlers and stopped making sense for apps that spend seconds waiting on a model. Web Functions brings Lambda close to a well-scaled container, without a cluster, and with revisions, canaries and multi-Region built in. The weak spots are the launch scope: Node.js only and five Regions. I'd try it now on a non-critical Node.js service, especially one that talks to Bedrock, and wait for the runtime and Region roadmap before making it a platform standard.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published in Spanish on &lt;a href="https://blog.cloudacademy.ar/lambda-web-functions-servidor-http-en-lambda-cpu-activa/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;, with the full CLI walkthrough, VPC connector setup and SCP.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/10/aws-lambda-web-functions/" rel="noopener noreferrer"&gt;What's New: AWS Lambda announces Lambda Web Functions&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/aws/announcing-aws-lambda-web-functions-for-fast-consistent-web-app-performance-anywhere/" rel="noopener noreferrer"&gt;AWS News Blog announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/10/cloudfront-lambda-web-functions-origin/" rel="noopener noreferrer"&gt;CloudFront supports Lambda Web Functions as an origin&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/lambda/latest/dg/web-functions.html" rel="noopener noreferrer"&gt;Lambda Web Functions developer guide&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/lambda/latest/dg/web-functions-scaling.html" rel="noopener noreferrer"&gt;Scaling behavior and quotas&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/lambda/latest/dg/web-functions-endpoint-access-control.html" rel="noopener noreferrer"&gt;Endpoint access control&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/lambda/pricing/" rel="noopener noreferrer"&gt;AWS Lambda pricing&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://blog.localstack.cloud/aws-lambda-web-functions-are-here/" rel="noopener noreferrer"&gt;LocalStack: test Web Functions locally&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>serverless</category>
      <category>lambda</category>
      <category>node</category>
    </item>
    <item>
      <title>AWS Well-Architected Agent: Your Architecture Review, Now Done by an Agent (Preview)</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Mon, 05 Oct 2026 12:55:30 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/aws-well-architected-agent-your-architecture-review-now-done-by-an-agent-preview-4k3m</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/aws-well-architected-agent-your-architecture-review-now-done-by-an-agent-preview-4k3m</guid>
      <description>&lt;p&gt;A proper Well-Architected review has always had the same cost: a day or two of the team answering a long questionnaire, someone pulling evidence from five consoles, and a report that's stale within weeks. Trusted Advisor helps with point checks, but it has no idea which application matters or what your business is optimizing for this quarter.&lt;/p&gt;

&lt;p&gt;On October 1, 2026, AWS launched &lt;strong&gt;AWS Well-Architected Agent&lt;/strong&gt; in public preview. It continuously analyzes your accounts and returns cost, security, resilience, and performance recommendations ranked against goals you write yourself, and it hands you the fix: console steps, CLI commands, or the change to your infrastructure code. AWS positions it as the next step after Trusted Advisor and the Well-Architected Tool, delivered through AWS Support.&lt;/p&gt;

&lt;p&gt;Here's how it works, how to wire it up with read-only cross-account roles, how to point it at an EKS Terraform repo before you deploy, and what to check before granting it access to your whole organization.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm9njki004anasai8zbk1.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm9njki004anasai8zbk1.webp" alt="How AWS Well-Architected Agent works: profile, read-only access, analysis, recommendations" width="800" height="390"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Profile → read-only role chaining → analysis → recommendations with remediation. (Diagram labels in Spanish, from the original article.)&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Public preview since Oct 1, 2026. The service runs in us-east-1, us-east-2, and us-west-2, but can analyze resources in &lt;strong&gt;any commercial Region&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;You create a &lt;strong&gt;profile&lt;/strong&gt;: up to 100 accounts, their Regions, the pillars you care about, and a goal per pillar. First recommendations land within 24 hours.&lt;/li&gt;
&lt;li&gt;It reads resource configuration across 65+ services, utilization metrics, and application topology via &lt;strong&gt;read-only&lt;/strong&gt; roles.&lt;/li&gt;
&lt;li&gt;Recommendations come at three levels (resource, application, architecture) with priority, effort, impact, ROI, and explicit trade-offs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;IaC review&lt;/strong&gt;: upload a .zip of Terraform, CloudFormation, or CDK and get findings before anything is deployed.&lt;/li&gt;
&lt;li&gt;Remediation via console, updated IaC, CLI, SDK, SSM runbooks, and MCP. In what AWS showed, it doesn't change resources on its own.&lt;/li&gt;
&lt;li&gt;Requires an AWS Support plan. No published preview pricing.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  What it is (and isn't)
&lt;/h2&gt;

&lt;p&gt;It's an AWS Support service surfaced in the Well-Architected console. It doesn't replace the Well-Architected Tool (AWS says you can keep using it for manual reviews), it isn't a compliance auditor, and every recommendation carries a disclaimer that it's AI-generated and may be wrong or incomplete.&lt;/p&gt;

&lt;p&gt;Preview scope: the console offers four pillars (cost, security, resilience, performance). The API already lists &lt;code&gt;OPERATIONAL_EXCELLENCE&lt;/code&gt; as a value, but it isn't exposed in the preview console. IaC reviews currently use the Well-Architected Framework lens.&lt;/p&gt;
&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Profile.&lt;/strong&gt; Created in a host account. You list target accounts (up to 100) and Regions, choose pillars, write a goal for each ("reduce unnecessary spend across my accounts", "workloads recover gracefully from disruptions"), and pick the execution role. Deletion protection is on by default in the console.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F7vo5acemq497ts7ljr2e.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F7vo5acemq497ts7ljr2e.webp" alt="Creating a Well-Architected Agent profile in the console" width="800" height="1200"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Profile setup: Regions, accounts, pillars with goals, roles. Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Application context&lt;/strong&gt; (optional). Tell it which applications exist, where they live, and which tags identify them. That's what turns "15 queues without alarms" into "your event pipeline has no failure observability in any of its three Regions."&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Access.&lt;/strong&gt; Role chaining: an execution role in the profile account trusts &lt;code&gt;wellarchitected.amazonaws.com&lt;/code&gt; and can assume an access role in each target account; access roles use the managed policy &lt;code&gt;WellArchitectedAgentResourceScanning&lt;/code&gt; and trust the execution role, with the profile ARN as external ID to prevent confused-deputy issues. Onboarding is opt-in per account.&lt;/p&gt;
&lt;h2&gt;
  
  
  Recommendations and remediation
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Type&lt;/th&gt;
&lt;th&gt;Scope&lt;/th&gt;
&lt;th&gt;AWS example&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Resource&lt;/td&gt;
&lt;td&gt;One resource or a small group&lt;/td&gt;
&lt;td&gt;CloudFormation role with &lt;code&gt;Action: "*"&lt;/code&gt; on &lt;code&gt;Resource: "*"&lt;/code&gt; that only needs 7 service namespaces&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Application (beta)&lt;/td&gt;
&lt;td&gt;Findings across one app&lt;/td&gt;
&lt;td&gt;15 DLQs in three Regions with no alarms; 48-minute median time to detect&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Architecture&lt;/td&gt;
&lt;td&gt;Patterns and IaC changes&lt;/td&gt;
&lt;td&gt;ECS Fargate tasks ~2x over-provisioned based on 30-day p95 CPU/memory&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Each comes with priority, effort, impact, an ROI estimate, affected resources, cross-pillar benefits, and &lt;strong&gt;trade-offs&lt;/strong&gt;. That last field is what a good architect always tells you, and here it's explicit.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5dyyj9cug96ohz1g1rrk.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5dyyj9cug96ohz1g1rrk.webp" alt="Well-Architected Agent dashboard with prioritized recommendations" width="800" height="930"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Recommendations ranked against the profile's goals. Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Start remediation&lt;/strong&gt; lets you pick console, updated IaC template, or AWS CLI. The API's remediation types also include SDK, MCP, and &lt;code&gt;AUTO_REMEDIATION&lt;/code&gt;. In AWS's example, the IaC option returns a CDK helper ready to paste into your stack, split into phases.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frjbtmm9gtmo8wilzoxho.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frjbtmm9gtmo8wilzoxho.webp" alt="Remediation options with updated IaC (CDK) code" width="799" height="648"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;IaC remediation returns the change for your repo. Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Architecture review.&lt;/strong&gt; Upload a .zip of your IaC project to S3, choose lens and pillars, start the review. Binary and media files are excluded. Classmethod tried it on day one: the review took about 30 minutes and flagged, among other things, a Lambda function hard-coded to &lt;code&gt;MemorySize: 256&lt;/code&gt; with no alarms or tuning automation.&lt;/p&gt;
&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;

&lt;p&gt;You need an active AWS Support plan, a recent AWS CLI, and one of the service Regions (us-east-1 here).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Execution role&lt;/strong&gt; (profile account):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cat&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; exec-trust.json &lt;span class="o"&gt;&amp;lt;&amp;lt;&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="no"&gt;EOF&lt;/span&gt;&lt;span class="sh"&gt;'
{"Version":"2012-10-17","Statement":[{"Effect":"Allow",
 "Principal":{"Service":"wellarchitected.amazonaws.com"},"Action":"sts:AssumeRole"}]}
&lt;/span&gt;&lt;span class="no"&gt;EOF
&lt;/span&gt;aws iam create-role &lt;span class="nt"&gt;--path&lt;/span&gt; /service-role/ &lt;span class="nt"&gt;--role-name&lt;/span&gt; WAAgentExecutionRole &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--assume-role-policy-document&lt;/span&gt; file://exec-trust.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Give it &lt;code&gt;sts:AssumeRole&lt;/code&gt; on &lt;code&gt;arn:aws:iam::*:role/AccessRoleForWellArchitectedAgent&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Access role&lt;/strong&gt; (each target account). Look up the managed policy ARN instead of hard-coding it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;POLICY_ARN&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;aws iam list-policies &lt;span class="nt"&gt;--scope&lt;/span&gt; AWS &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--query&lt;/span&gt; &lt;span class="s2"&gt;"Policies[?PolicyName=='WellArchitectedAgentResourceScanning'].Arn"&lt;/span&gt; &lt;span class="nt"&gt;--output&lt;/span&gt; text&lt;span class="si"&gt;)&lt;/span&gt;
aws iam create-role &lt;span class="nt"&gt;--role-name&lt;/span&gt; AccessRoleForWellArchitectedAgent &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--assume-role-policy-document&lt;/span&gt; file://access-trust.json
aws iam attach-role-policy &lt;span class="nt"&gt;--role-name&lt;/span&gt; AccessRoleForWellArchitectedAgent &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--policy-arn&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$POLICY_ARN&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;At scale, deploy this with CloudFormation StackSets, and add the external ID condition per the IAM docs once the profile exists.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Profile:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws wellarchitected create-agent-profile &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--name&lt;/span&gt; platform-eks &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--business-overview&lt;/span&gt; &lt;span class="s2"&gt;"EKS container platform for payment APIs"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--pillars&lt;/span&gt; COST_OPTIMIZATION SECURITY RESILIENCE PERFORMANCE &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--execution-role-arn&lt;/span&gt; arn:aws:iam::111122223333:role/service-role/WAAgentExecutionRole &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--aggregation-configuration&lt;/span&gt; file://aggregation.json &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--deletion-protection&lt;/span&gt; &lt;span class="nt"&gt;--region&lt;/span&gt; us-east-1
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;aggregation.json&lt;/code&gt; is a list of &lt;code&gt;{"accountId", "regions", "accessRoleArn"}&lt;/code&gt; objects. &lt;code&gt;get-agent-profile&lt;/code&gt; reports &lt;code&gt;eligibleForScheduledGeneration&lt;/code&gt;, &lt;code&gt;eligibleForArchitectureGeneration&lt;/code&gt;, and any &lt;code&gt;fieldErrors&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Review an EKS Terraform repo&lt;/strong&gt; — package it without state, provider cache, or secrets:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;zip &lt;span class="nt"&gt;-r&lt;/span&gt; ../infra-eks.zip &lt;span class="nb"&gt;.&lt;/span&gt; &lt;span class="nt"&gt;-x&lt;/span&gt; &lt;span class="s2"&gt;".terraform/*"&lt;/span&gt; &lt;span class="s2"&gt;"*.tfstate"&lt;/span&gt; &lt;span class="s2"&gt;"*.tfstate.backup"&lt;/span&gt; &lt;span class="s2"&gt;"*.tfvars"&lt;/span&gt; &lt;span class="s2"&gt;".git/*"&lt;/span&gt;
aws s3 &lt;span class="nb"&gt;cp&lt;/span&gt; ../infra-eks.zip s3://my-review-bucket/infra-eks.zip
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then &lt;strong&gt;Conduct architecture review&lt;/strong&gt; in the console.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Pull recommendations from the CLI:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws wellarchitected list-agent-recommendations &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--profile-arn&lt;/span&gt; arn:aws:wellarchitected:us-east-1:111122223333:agent-profile/platform-eks &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--pillar&lt;/span&gt; COST_OPTIMIZATION &lt;span class="nt"&gt;--state&lt;/span&gt; OPEN &lt;span class="nt"&gt;--region&lt;/span&gt; us-east-1

aws wellarchitected get-agent-recommendation &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--recommendation-arn&lt;/span&gt; &amp;lt;recommendation ARN&amp;gt; &lt;span class="nt"&gt;--remediation-type&lt;/span&gt; IAC &lt;span class="nt"&gt;--region&lt;/span&gt; us-east-1
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Each recommendation includes &lt;code&gt;awsServices&lt;/code&gt;, so you can filter EKS/ECS items into a weekly platform report or open an issue per &lt;code&gt;HIGH&lt;/code&gt; finding.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Traps:&lt;/strong&gt; applying CLI fixes in a GitOps setup (Terraform or Argo CD will revert the drift; route IaC fixes through PRs); zipping &lt;code&gt;tfvars&lt;/code&gt; or &lt;code&gt;.env&lt;/code&gt; files; assuming read-only means nothing sensitive is read; expecting results before 24 hours; treating dollar figures as exact; forgetting deletion protection when you try to remove a profile.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pricing and availability
&lt;/h2&gt;

&lt;p&gt;No published price during preview; it requires an active AWS Support plan. You pay normally for the S3 object holding your IaC zip and for whatever remediation deploys (alarms, SSM runbooks, capacity changes).&lt;/p&gt;

&lt;p&gt;The service runs only in three US Regions but can onboard resources from any commercial Region. That's the part to check: if your workloads live in, say, São Paulo, their configuration, metrics, and uploaded IaC get processed in the US, and the announcement doesn't say where that data is stored or for how long. Regulated workloads should get a compliance sign-off first. And it's a preview, so APIs, pillars, lenses, and pricing can change.&lt;/p&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;What matters here isn't that it uses AI. It's that the unit of work changes. Trusted Advisor gave you checks; the Well-Architected Tool gave you a questionnaire. This gives you a recommendation with context, impact, trade-offs, and the change for your repo, ranked by goals you wrote. If quality holds up in messy real environments (AWS's demo was very polished; Classmethod got reasonable but medium/low-priority findings), it'll replace most annual manual reviews. Two things are still unknown: price and data residency. I'd try it now on non-production accounts, route every fix through a pull request, and wait for GA pricing before rolling it out org-wide.&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/aws/announcing-aws-well-architected-agent-an-ai-powered-intelligence-to-optimize-your-cloud-environment-preview/" rel="noopener noreferrer"&gt;AWS News Blog announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/10/aws-well-architected-agent/" rel="noopener noreferrer"&gt;What's New&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://docs.aws.amazon.com/wellarchitected/latest/userguide/wa-agent.html" rel="noopener noreferrer"&gt;User guide&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/wellarchitected/latest/userguide/security-iam-agent.html" rel="noopener noreferrer"&gt;IAM for the agent&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;CLI: &lt;a href="https://docs.aws.amazon.com/cli/latest/reference/wellarchitected/create-agent-profile.html" rel="noopener noreferrer"&gt;create-agent-profile&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/cli/latest/reference/wellarchitected/list-agent-recommendations.html" rel="noopener noreferrer"&gt;list-agent-recommendations&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/cli/latest/reference/wellarchitected/get-agent-recommendation.html" rel="noopener noreferrer"&gt;get-agent-recommendation&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;&lt;a href="https://dev.classmethod.jp/articles/aws-wa-agent-preview/" rel="noopener noreferrer"&gt;Classmethod hands-on (Japanese)&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Originally published in Spanish on &lt;a href="https://blog.cloudacademy.ar/aws-well-architected-agent-preview/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>aws</category>
      <category>devops</category>
      <category>ai</category>
      <category>cloud</category>
    </item>
    <item>
      <title>S3 Vectors Now Filters Before It Searches: Metadata Pre-Filtering for Multi-Tenant RAG</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Mon, 05 Oct 2026 12:55:30 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/s3-vectors-now-filters-before-it-searches-metadata-pre-filtering-for-multi-tenant-rag-5eaj</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/s3-vectors-now-filters-before-it-searches-metadata-pre-filtering-for-multi-tenant-rag-5eaj</guid>
      <description>&lt;p&gt;If you run RAG on Amazon S3 Vectors with metadata filters, you may have been losing results without knowing it. Ask for the 10 nearest chunks &lt;em&gt;for one tenant&lt;/em&gt; and you might get 3 back, or 10 that aren't actually the closest ones in that tenant's data. No error, no alarm, just a model answering with less context than it should have had.&lt;/p&gt;

&lt;p&gt;On September 30, 2026, AWS shipped &lt;strong&gt;metadata pre-filtering&lt;/strong&gt; for S3 Vectors. Indexes now come in two modes, and the new one evaluates your filter &lt;em&gt;before&lt;/em&gt; the similarity search instead of during it. AWS says that returns up to 5x more matching vectors when the filter is selective. It also adds a &lt;code&gt;$startsWith&lt;/code&gt; prefix operator, sets a 100-condition limit per query, and costs nothing extra.&lt;/p&gt;

&lt;p&gt;Here's what changed, why the old behavior dropped results, how to measure the difference on your own index, and what a multi-tenant retrieval service on Amazon EKS looks like with it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbcggf908p08c8landog5.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbcggf908p08c8landog5.webp" alt="Diagram comparing CLASSIC and ENHANCED query processing in S3 Vectors" width="800" height="350"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;CLASSIC validates the filter while searching; ENHANCED narrows the set first. (Diagram labels in Spanish, from the original article.)&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;S3 Vectors indexes now have a mode: &lt;code&gt;CLASSIC&lt;/code&gt; (filter evaluated in tandem with the approximate search) or &lt;code&gt;ENHANCED&lt;/code&gt; (filter resolved first, then similarity search over the matching subset).&lt;/li&gt;
&lt;li&gt;With selective filters, &lt;code&gt;ENHANCED&lt;/code&gt; returns up to 5x more matching vectors, per AWS.&lt;/li&gt;
&lt;li&gt;Vector buckets created on or after Sep 30, 2026 are &lt;code&gt;ENHANCED&lt;/code&gt;-only. Older buckets stay &lt;code&gt;CLASSIC&lt;/code&gt; until you run &lt;code&gt;UpdateIndexMode&lt;/code&gt;; no re-ingestion needed.&lt;/li&gt;
&lt;li&gt;You can test per query with &lt;code&gt;queryMode=ENHANCED&lt;/code&gt; before switching the index.&lt;/li&gt;
&lt;li&gt;New &lt;code&gt;$startsWith&lt;/code&gt; operator; &lt;code&gt;ENHANCED&lt;/code&gt; indexes cap filters at 100 conditions (each &lt;code&gt;$in&lt;/code&gt; value counts).&lt;/li&gt;
&lt;li&gt;No additional cost. Available in every commercial Region with S3 Vectors, including São Paulo, plus the China Regions.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  Why the old mode lost results
&lt;/h2&gt;

&lt;p&gt;Vector search at scale is approximate: the engine explores a neighborhood of promising candidates on a bounded budget rather than comparing your query against every vector.&lt;/p&gt;

&lt;p&gt;In &lt;code&gt;CLASSIC&lt;/code&gt; mode, filters are checked &lt;em&gt;while&lt;/em&gt; that exploration runs. If your filter matches half the corpus, most candidates pass and nothing bad happens. If it matches 0.005%, almost every candidate gets thrown away, the budget runs out, and you get fewer than K results, or K results that aren't the true nearest neighbors within the filtered set. The docs say it plainly: in &lt;code&gt;CLASSIC&lt;/code&gt;, filtered queries may return fewer than K results when few vectors match.&lt;/p&gt;

&lt;p&gt;An AWS re:Post article measured this on one million synthetic vectors: recall@10 went from 0.86 unfiltered to 0.83 at 10% selectivity, 0.42 at 1%, and 0.15 at 0.01%. The workaround was to partition data into separate indexes by low-cardinality fields (tenant, language, Region), which recovered 15–31 recall points at the cost of managing more indexes.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;ENHANCED&lt;/code&gt; flips the order. AWS's example: 8 million support tickets, one customer owns 400. With pre-filtering, the similarity search runs over those 400, not over candidates drawn from all 8 million.&lt;/p&gt;

&lt;p&gt;One takeaway worth repeating: &lt;strong&gt;result count is not a quality signal&lt;/strong&gt;. Getting 10 results back in &lt;code&gt;CLASSIC&lt;/code&gt; never guaranteed they were the best 10.&lt;/p&gt;
&lt;h2&gt;
  
  
  Index modes at a glance
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Situation&lt;/th&gt;
&lt;th&gt;Default mode&lt;/th&gt;
&lt;th&gt;Can you change it?&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Buckets created on/after Sep 30, 2026&lt;/td&gt;
&lt;td&gt;&lt;code&gt;ENHANCED&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;No way back to &lt;code&gt;CLASSIC&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Buckets created before Sep 30, 2026&lt;/td&gt;
&lt;td&gt;&lt;code&gt;CLASSIC&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Yes: to &lt;code&gt;ENHANCED&lt;/code&gt; via &lt;code&gt;UpdateIndexMode&lt;/code&gt;, and back to &lt;code&gt;CLASSIC&lt;/code&gt; via CLI/SDK/API only&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;New indexes in an older bucket&lt;/td&gt;
&lt;td&gt;Bucket default&lt;/td&gt;
&lt;td&gt;Set with &lt;code&gt;PutVectorBucketDefaultIndexMode&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;code&gt;UpdateIndexMode&lt;/code&gt; affects only that index. The console offers &lt;strong&gt;Enable enhanced index mode&lt;/strong&gt; on the index detail page; reverting is not available in the console. The current mode shows up as &lt;code&gt;indexMode&lt;/code&gt; in &lt;code&gt;GetIndex&lt;/code&gt;.&lt;/p&gt;
&lt;h2&gt;
  
  
  Filters: operators and the 100-condition limit
&lt;/h2&gt;

&lt;p&gt;Filter syntax is unchanged: &lt;code&gt;$eq&lt;/code&gt;, &lt;code&gt;$ne&lt;/code&gt;, &lt;code&gt;$gt&lt;/code&gt;, &lt;code&gt;$gte&lt;/code&gt;, &lt;code&gt;$lt&lt;/code&gt;, &lt;code&gt;$lte&lt;/code&gt;, &lt;code&gt;$in&lt;/code&gt;, &lt;code&gt;$nin&lt;/code&gt;, &lt;code&gt;$exists&lt;/code&gt;, &lt;code&gt;$and&lt;/code&gt;, &lt;code&gt;$or&lt;/code&gt;, plus the new &lt;code&gt;$startsWith&lt;/code&gt; (string prefix, &lt;code&gt;ENHANCED&lt;/code&gt; only). Prefix matching is handy for path-organized collections: &lt;code&gt;{"s3_path": {"$startsWith": "matter-4417/exhibits/"}}&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The 100-condition cap applies to &lt;code&gt;ENHANCED&lt;/code&gt; indexes and is counted per value: &lt;code&gt;{"region": {"$in": ["us-east-1", "us-west-2", "eu-west-1"]}}&lt;/code&gt; is 3 conditions. If you build dynamic filters with long &lt;code&gt;$in&lt;/code&gt; lists (say, every project a user can access), audit them before migrating: over 100, an &lt;code&gt;ENHANCED&lt;/code&gt; index returns a validation error.&lt;/p&gt;

&lt;p&gt;Metadata limits haven't changed: up to 40 KB per vector, of which up to 2 KB is filterable; up to 50 keys; up to 10 non-filterable keys per index, fixed at index creation. Keep filter fields short and put chunk text in non-filterable metadata. Bedrock Knowledge Bases already does this with &lt;code&gt;AMAZON_BEDROCK_TEXT&lt;/code&gt; and &lt;code&gt;AMAZON_BEDROCK_METADATA&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fq2u6skyi8rux3zrea385.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fq2u6skyi8rux3zrea385.webp" alt="S3 console index creation form showing non-filterable metadata keys" width="800" height="508"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Non-filterable keys are declared at index creation and can't be changed later. Image: AWS.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;

&lt;p&gt;You'll need a recent AWS CLI and boto3.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Check the mode&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws s3vectors get-index &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--vector-bucket-name&lt;/span&gt; my-vector-bucket &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--index-name&lt;/span&gt; tickets &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--query&lt;/span&gt; &lt;span class="s1"&gt;'index.indexMode'&lt;/span&gt; &lt;span class="nt"&gt;--output&lt;/span&gt; text
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;2. Measure recall before migrating.&lt;/strong&gt; Pick a small tenant, compute its true top-K by brute force, and compare against the index in each mode. This reads the whole index with &lt;code&gt;ListVectors&lt;/code&gt;, so do it on a test copy or a sample.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;boto3&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;numpy&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;

&lt;span class="n"&gt;BUCKET&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;INDEX&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;TENANT&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;K&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;my-vector-bucket&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;tickets&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;t-10428&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;10&lt;/span&gt;
&lt;span class="n"&gt;s3v&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;boto3&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;client&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;s3vectors&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;bedrock&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;boto3&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;client&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;bedrock-runtime&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;embed&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;body&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;dumps&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;inputText&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;dimensions&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;1024&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;normalize&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;})&lt;/span&gt;
    &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;bedrock&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;invoke_model&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;modelId&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;amazon.titan-embed-text-v2:0&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;loads&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;body&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="nf"&gt;read&lt;/span&gt;&lt;span class="p"&gt;())[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;embedding&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;

&lt;span class="n"&gt;keys&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;mat&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;token&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[],&lt;/span&gt; &lt;span class="p"&gt;[],&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;
&lt;span class="k"&gt;while&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;kw&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;dict&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;vectorBucketName&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;BUCKET&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;indexName&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;INDEX&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;maxResults&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
              &lt;span class="n"&gt;returnData&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;returnMetadata&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;token&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;kw&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;nextToken&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;token&lt;/span&gt;
    &lt;span class="n"&gt;page&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;s3v&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;list_vectors&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;kw&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;v&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;page&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;vectors&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]:&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;v&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;metadata&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{}).&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;tenant_id&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;TENANT&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="n"&gt;keys&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;v&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;key&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]);&lt;/span&gt; &lt;span class="n"&gt;mat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;v&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;data&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;float32&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
    &lt;span class="n"&gt;token&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;page&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;nextToken&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;token&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;break&lt;/span&gt;
&lt;span class="n"&gt;mat&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;array&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;mat&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;ground_truth&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;q&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;array&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;sims&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;mat&lt;/span&gt; &lt;span class="o"&gt;@&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;linalg&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;norm&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;mat&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;axis&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;linalg&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;norm&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="n"&gt;keys&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;argsort&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;sims&lt;/span&gt;&lt;span class="p"&gt;)[:&lt;/span&gt;&lt;span class="n"&gt;K&lt;/span&gt;&lt;span class="p"&gt;]}&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;query&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;mode&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;s3v&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;query_vectors&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;vectorBucketName&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;BUCKET&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;indexName&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;INDEX&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                          &lt;span class="n"&gt;queryVector&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;float32&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt; &lt;span class="n"&gt;topK&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;K&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                          &lt;span class="nb"&gt;filter&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;tenant_id&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;TENANT&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt; &lt;span class="n"&gt;queryMode&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;mode&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="n"&gt;v&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;key&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;v&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;vectors&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]}&lt;/span&gt;

&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;can&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;t log in&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;billing error&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;plan change&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]:&lt;/span&gt;
    &lt;span class="n"&gt;q&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;embed&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt; &lt;span class="n"&gt;gt&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;ground_truth&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;mode&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;CLASSIC&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;ENHANCED&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="n"&gt;got&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;query&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;mode&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;mode&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;9&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;14&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; returned=&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;got&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt; recall@&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;K&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;=&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;got&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;gt&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;gt&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This assumes cosine distance and 1024-dim Titan Text Embeddings V2; adjust to your index.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Migrate&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws s3vectors update-index-mode &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--vector-bucket-name&lt;/span&gt; my-vector-bucket &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--index-name&lt;/span&gt; tickets &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--index-mode&lt;/span&gt; ENHANCED

aws s3vectors put-vector-bucket-default-index-mode &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--vector-bucket-name&lt;/span&gt; my-vector-bucket &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--default-index-mode&lt;/span&gt; ENHANCED
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;4. Retrieval on EKS with Pod Identity.&lt;/strong&gt; The tenant ID comes from the user's token, never from a client-controlled parameter. The API builds the filter; the pod queries S3 Vectors with a role scoped to one index.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqq638aoz6wxjz22udgad.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqq638aoz6wxjz22udgad.webp" alt="Multi-tenant retrieval architecture on Amazon EKS querying an ENHANCED S3 Vectors index" width="800" height="380"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;The tenant filter is built server-side and resolved before the similarity search. (Diagram labels in Spanish.)&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;A permissions gotcha: &lt;code&gt;s3vectors:QueryVectors&lt;/code&gt; alone only works without filters and without &lt;code&gt;returnMetadata&lt;/code&gt;. With either, you also need &lt;code&gt;s3vectors:GetVectors&lt;/code&gt;, or you get a 403.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"Version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"2012-10-17"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"Statement"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"Effect"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Allow"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"Action"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"s3vectors:QueryVectors"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"s3vectors:GetVectors"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"Resource"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"arn:aws:s3vectors:sa-east-1:111122223333:bucket/my-vector-bucket/index/tickets"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}]&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Trust policy principal is &lt;code&gt;pods.eks.amazonaws.com&lt;/code&gt; with &lt;code&gt;sts:AssumeRole&lt;/code&gt; and &lt;code&gt;sts:TagSession&lt;/code&gt;, and the cluster needs the &lt;code&gt;eks-pod-identity-agent&lt;/code&gt; add-on. Then:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws eks create-pod-identity-association &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--cluster-name&lt;/span&gt; my-cluster &lt;span class="nt"&gt;--namespace&lt;/span&gt; rag &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--service-account&lt;/span&gt; retrieval &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--role-arn&lt;/span&gt; arn:aws:iam::111122223333:role/rag-retrieval
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And the search function:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;boto3&lt;/span&gt;
&lt;span class="n"&gt;s3v&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;boto3&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;client&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;s3vectors&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;search&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;tenant_id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;query_embedding&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;folder&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;None&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;k&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;8&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;conds&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;tenant_id&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;tenant_id&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;active&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;}]&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;folder&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;conds&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;s3_path&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;$startsWith&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;folder&lt;/span&gt;&lt;span class="p"&gt;}})&lt;/span&gt;
    &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;s3v&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;query_vectors&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;vectorBucketName&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;VECTOR_BUCKET&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
        &lt;span class="n"&gt;indexName&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;VECTOR_INDEX&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
        &lt;span class="n"&gt;queryVector&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;float32&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;query_embedding&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="n"&gt;topK&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;k&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;filter&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;$and&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;conds&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="n"&gt;returnMetadata&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;returnDistance&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;vectors&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Common traps:&lt;/strong&gt; treating a tenant filter as isolation (a bug that drops it exposes the whole index; use per-tenant indexes with separate IAM for strict isolation); exceeding 100 conditions; using &lt;code&gt;$startsWith&lt;/code&gt; on a &lt;code&gt;CLASSIC&lt;/code&gt; index without &lt;code&gt;queryMode=ENHANCED&lt;/code&gt;; storing dates as strings (range operators are numeric only); and skipping latency tests with broad filters, since AWS notes latency grows with index size, filter breadth, and condition count.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;No extra charge for pre-filtering. S3 Vectors pricing in us-east-1: storage $0.06/GB-month; PUT $0.20/GB uploaded; $2.50 per million query API calls; plus per-TB data-processed charges tiered by index size, and the first 512 KB returned per query free. AWS's published example (10M 1024-dim vectors across 40 indexes, 1M queries/month at top-100) comes to &lt;strong&gt;$11.38/month&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Note from that example: processing is driven by the size of the index you query, not by how many vectors pass the filter. A selective filter improves recall but doesn't make the query cheaper. Large tenants that always query their own data are still good candidates for a dedicated index.&lt;/p&gt;

&lt;h2&gt;
  
  
  Availability and caveats
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Regions:&lt;/strong&gt; all commercial Regions where S3 Vectors runs (including sa-east-1) and the China Regions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Data residency:&lt;/strong&gt; vectors stay in the bucket's Region; check where your embedding model runs, especially with cross-Region inference profiles.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cross-Region traffic:&lt;/strong&gt; keep your compute and your vector bucket in the same Region.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bedrock Knowledge Bases:&lt;/strong&gt; the announcement doesn't detail how &lt;code&gt;Retrieve&lt;/code&gt; filters interact with index mode. If your knowledge base sits on a pre-Sep 30 bucket, check the index mode and measure.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;This fixes a limitation that forced you to design around the engine instead of the problem. Until now, the honest answer to "can all my tenants share one index?" was "yes, but small tenants will lose results." With pre-filtering, a shared index becomes reasonable for the long tail, while large tenants or those needing strict IAM isolation keep dedicated indexes. What doesn't change: the filter is still your code, and in multi-tenant RAG a missing filter is a data leak. I'd run the recall test above, migrate &lt;code&gt;CLASSIC&lt;/code&gt; indexes, and use the occasion to audit where tenant filters get built.&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/09/s3-vectors-introduces-metadata-pre-filtering/" rel="noopener noreferrer"&gt;What's New: S3 Vectors introduces metadata pre-filtering&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/aws/amazon-s3-vectors-now-supports-metadata-pre-filtering-for-higher-recall-on-filtered-searches/" rel="noopener noreferrer"&gt;AWS News Blog announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://docs.aws.amazon.com/AmazonS3/latest/userguide/s3-vectors-metadata-filtering.html" rel="noopener noreferrer"&gt;Docs: Metadata filtering&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/AmazonS3/latest/userguide/s3-vectors-index-mode.html" rel="noopener noreferrer"&gt;Changing index mode&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/AmazonS3/latest/userguide/s3-vectors-limitations.html" rel="noopener noreferrer"&gt;Limits&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/s3/pricing/" rel="noopener noreferrer"&gt;S3 pricing (S3 Vectors tab)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://repost.aws/articles/ARpXgjTc00RZG4mfUYQ7Y-QQ/how-metadata-filter-selectivity-affects-recall-in-amazon-s3-vectors" rel="noopener noreferrer"&gt;re:Post: How metadata filter selectivity affects recall in S3 Vectors&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Originally published in Spanish on &lt;a href="https://blog.cloudacademy.ar/s3-vectors-pre-filtrado-de-metadata-rag-multi-tenant/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>aws</category>
      <category>rag</category>
      <category>kubernetes</category>
      <category>ai</category>
    </item>
    <item>
      <title>EventBridge's new custom event bus: one bus for the whole org, with ordering, replay and per-GB pricing</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Mon, 28 Sep 2026 15:21:00 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/eventbridges-new-custom-event-bus-one-bus-for-the-whole-org-with-ordering-replay-and-per-gb-3k2g</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/eventbridges-new-custom-event-bus-one-bus-for-the-whole-org-with-ordering-replay-and-per-gb-3k2g</guid>
      <description>&lt;p&gt;On September 24, 2026, AWS relaunched the Amazon EventBridge custom event bus. This isn't a new toggle on the bus you already know. It's a separate resource with its own API namespace (&lt;code&gt;eventsv2&lt;/code&gt;), its own sharing model, and a pricing model based on data volume instead of event count. Your existing buses are renamed "Custom event bus – classic" and keep working exactly as before.&lt;/p&gt;

&lt;p&gt;The interesting part is what AWS is going after. EventBridge has always been pleasant inside a single account, but in a multi-account organization you end up with a mesh of buses, forwarding rules and resource policies, and the platform team becomes the ticket queue for every new subscription. And the moment someone needs strict ordering or wants to reprocess last Tuesday's events, the answer used to be "put SQS FIFO, Kinesis or Kafka in the middle."&lt;/p&gt;

&lt;p&gt;The new bus tackles both problems: one bus shared across the organization, per-group FIFO delivery, deduplication, retention of up to a year, and replay from any point in time.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;New &lt;strong&gt;custom event bus&lt;/strong&gt; resource (&lt;code&gt;eventsv2&lt;/code&gt; API). The old one becomes &lt;em&gt;classic&lt;/em&gt; and doesn't change.&lt;/li&gt;
&lt;li&gt;Share one bus with accounts, OUs or the whole organization through &lt;strong&gt;AWS RAM&lt;/strong&gt; using four managed permissions. Default quota: 10,000 subscribers per bus.&lt;/li&gt;
&lt;li&gt;New &lt;strong&gt;Subscriber&lt;/strong&gt; resource: filter + transformation + one target + retries + DLQ, owned by each consuming account.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;FIFO per &lt;code&gt;EventGroupId&lt;/code&gt;&lt;/strong&gt;, &lt;strong&gt;deduplication&lt;/strong&gt; within a 5-minute window, &lt;strong&gt;retention&lt;/strong&gt; from 1 to 365 days, &lt;strong&gt;replay&lt;/strong&gt; via each subscriber's starting position.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;PutRawEvents&lt;/code&gt; accepts JSON, Avro, Protobuf or raw bytes with custom metadata. Transformations are written in &lt;strong&gt;JSONata&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Pricing: $0.18/GB ingress ($0.12 after 5,000 GB), $0.05/GB egress per subscriber, $0.08/GB-month extra retention, $0.15 per million evaluations.&lt;/li&gt;
&lt;li&gt;14 Regions at launch. &lt;strong&gt;Not in São Paulo (sa-east-1)&lt;/strong&gt; yet.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F06z9tqqd09oe0y5ldq1e.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F06z9tqqd09oe0y5ldq1e.webp" alt="A shared bus in the platform account with publishing accounts on the left and team accounts owning their subscribers on the right" width="800" height="430"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;One bus in the platform account, shared via AWS RAM. Each team creates, and pays for, its own subscribers.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  The ownership model is the real change
&lt;/h2&gt;

&lt;p&gt;The classic model has the bus owner writing rules on behalf of everyone. In the new model, the platform team creates the bus and shares it; each team creates subscribers in its own account, only sees its own subscribers when listing, and can't modify or delete the bus. The owner sees every subscriber and can cut one off with &lt;code&gt;RevokeResource&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Sharing uses four RAM managed permissions: &lt;code&gt;AWSRAMEventBridgeEventBusV2PublishOnly&lt;/code&gt;, &lt;code&gt;...SubscribeOnly&lt;/code&gt;, &lt;code&gt;...EventSourceAccess&lt;/code&gt; and &lt;code&gt;...FullAccess&lt;/code&gt;. Two constraints to design around: consumers can't re-share the bus, and &lt;strong&gt;the bus and its subscribers must live in the same Region and partition&lt;/strong&gt;. Sharing is free, and the owner gets per-consumer &lt;code&gt;EventsDelivered&lt;/code&gt; and &lt;code&gt;EgressBytes&lt;/code&gt; metrics, which is handy for chargeback.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5sv780ono7mvkfxz7nsu.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5sv780ono7mvkfxz7nsu.webp" alt="EventBridge console showing the new and classic custom event bus options" width="800" height="630"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Creating a custom event bus: new vs. classic. Image: AWS.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxic2o00152msko18f4jg.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxic2o00152msko18f4jg.webp" alt="Sharing the bus with AWS Resource Access Manager" width="800" height="362"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Sharing the bus through AWS RAM. Image: AWS.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Subscribers
&lt;/h2&gt;

&lt;p&gt;A subscriber bundles what used to be spread across a rule, its targets, an input transformer, a retry policy and a DLQ:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Filters&lt;/strong&gt; use regular event pattern syntax with three scopes: &lt;code&gt;DATA&lt;/code&gt; (payload), &lt;code&gt;METADATA&lt;/code&gt; (producer-supplied key/value pairs, exact match only) and &lt;code&gt;SYSTEM_METADATA&lt;/code&gt; (fields EventBridge sets, such as &lt;code&gt;EventGroupId&lt;/code&gt;). An event must match every filter.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Transformers&lt;/strong&gt;: &lt;code&gt;RAW&lt;/code&gt; (default), &lt;code&gt;WITH_METADATA&lt;/code&gt;, or &lt;code&gt;JSONATA&lt;/code&gt; with the event exposed as &lt;code&gt;$events&lt;/code&gt;. Syntax is validated at creation, but evaluation happens at delivery time, so a bad expression fails deliveries, not deployments.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One target per subscriber&lt;/strong&gt;: SQS, Lambda, Kinesis, Step Functions, SNS, HTTP/API destinations, Firehose, another bus, or universal targets like &lt;code&gt;arn:aws:events:::aws-sdk:dynamodb:putItem&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Retries&lt;/strong&gt;: 5 attempts within 300 seconds by default, configurable up to 185 attempts and 24 hours, then an SQS DLQ.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmasn4k4aq0hnun0uosuh.webp" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmasn4k4aq0hnun0uosuh.webp" alt="Event lifecycle: publish, deduplicate, store, filter, transform, deliver, retry and DLQ" width="800" height="310"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;What happens to an event on the new bus, per subscriber.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  Ordering and deduplication, with the fine print
&lt;/h2&gt;

&lt;p&gt;Producers set &lt;code&gt;SystemMetadata.EventGroupId&lt;/code&gt;. Subscribers configured for ordered delivery get events from the same group in sequence; other subscribers get the same events unordered. For Lambda or Express state machines you can invoke synchronously (&lt;code&gt;REQUEST_RESPONSE&lt;/code&gt;) so EventBridge waits for the result before moving on. Standard state machines don't support that mode.&lt;/p&gt;

&lt;p&gt;The docs are upfront about the tradeoff: &lt;strong&gt;a failing event blocks later events in its group for as long as it's retried.&lt;/strong&gt; Keep the retry window short on ordered subscribers and use a FIFO DLQ, where EventBridge sets &lt;code&gt;MessageGroupId&lt;/code&gt; from the &lt;code&gt;EventGroupId&lt;/code&gt; so failures stay ordered too.&lt;/p&gt;

&lt;p&gt;Deduplication happens at publish time, within 5 minutes, using your &lt;code&gt;DeduplicationId&lt;/code&gt; or a content hash. Duplicates come back as &lt;code&gt;DEDUPLICATED&lt;/code&gt;, which is a success code. AWS frames this as exactly-once semantics; I'd read it more narrowly as protection against producer retries. Your consumers still need to be idempotent.&lt;/p&gt;
&lt;h2&gt;
  
  
  Hands-on
&lt;/h2&gt;

&lt;p&gt;Create a bus with 30 days of retention and a customer managed key:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws eventsv2 create-event-bus &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--name&lt;/span&gt; orders &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--description&lt;/span&gt; &lt;span class="s2"&gt;"Order events from the checkout service"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--storage-configuration&lt;/span&gt; &lt;span class="nv"&gt;RetentionPeriodInDays&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;30 &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--encryption-configuration&lt;/span&gt; &lt;span class="nv"&gt;KmsKeyIdentifier&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;arn:aws:kms:us-east-1:111122223333:key/1234abcd-12ab-34cd-56ef-1234567890ab &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--tags&lt;/span&gt; &lt;span class="nv"&gt;team&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;orders
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Wait for &lt;code&gt;ACTIVE&lt;/code&gt; before publishing. A KMS problem leaves it in &lt;code&gt;CREATE_FAILED&lt;/code&gt;, and all you can do is delete it.&lt;/p&gt;

&lt;p&gt;Share it with a consuming account:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws ram create-resource-share &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--name&lt;/span&gt; orders-bus-share &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--resource-arns&lt;/span&gt; arn:aws:events:us-east-1:111122223333:event-busv2/orders/EXAMPLE1234567890abcdef &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--principals&lt;/span&gt; 444455556666 &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--permission-arns&lt;/span&gt; arn:aws:ram::aws:permission/AWSRAMEventBridgeEventBusV2SubscribeOnly
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;From the consuming account, create a subscriber for large orders with a DLQ:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws eventsv2 create-subscriber &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--name&lt;/span&gt; large-orders &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--event-bus-arn&lt;/span&gt; arn:aws:events:us-east-1:111122223333:event-busv2/orders/EXAMPLE1234567890abcdef &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--filter-configuration&lt;/span&gt; &lt;span class="s1"&gt;'{
        "Filters": [
            { "Scope": "DATA", "Pattern": "{\"detail\":{\"total\":[{\"numeric\":[\"&amp;gt;\",500]}]}}" }
        ]
    }'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--invoke-configuration&lt;/span&gt; &lt;span class="s1"&gt;'{
        "TargetArn": "arn:aws:sqs:us-east-1:111122223333:large-orders",
        "RoleArn": "arn:aws:iam::111122223333:role/EventBusDeliveryRole"
    }'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
    &lt;span class="nt"&gt;--on-failure-configuration&lt;/span&gt; &lt;span class="s1"&gt;'{ "Arn": "arn:aws:sqs:us-east-1:111122223333:large-orders-dlq" }'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Reprocess from a timestamp by creating a subscriber with &lt;code&gt;--starting-position POINT_IN_TIME&lt;/code&gt; and &lt;code&gt;--point-in-time-configuration '{ "PointType": "TIMESTAMP", "StartingPoint": "2026-09-01T00:00:00Z" }'&lt;/code&gt;. Pause with &lt;code&gt;update-subscriber --state STOPPED&lt;/code&gt;, and resume with &lt;code&gt;--resume-position LAST_PROCESSED&lt;/code&gt; (drain the backlog) or &lt;code&gt;LATEST&lt;/code&gt; (skip it).&lt;/p&gt;

&lt;p&gt;Gotchas worth knowing before your first 2 a.m. page:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Only one subscriber create/delete runs at a time per bus. Parallel IaC gets &lt;code&gt;ConcurrentModificationException&lt;/code&gt;, so retry or serialize.&lt;/li&gt;
&lt;li&gt;Filters depend on the publish API. A pattern written for the &lt;code&gt;PutEvents&lt;/code&gt; envelope matches nothing when the producer uses &lt;code&gt;PutRawEvents&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;A &lt;code&gt;LATEST&lt;/code&gt; subscriber won't see events published before it existed.&lt;/li&gt;
&lt;li&gt;For FIFO targets, derive &lt;code&gt;MessageDeduplicationId&lt;/code&gt; per message with JSONata. A constant value means SQS drops everything after the first message.&lt;/li&gt;
&lt;li&gt;Debug with the &lt;code&gt;AWS/EventsV2&lt;/code&gt; metrics as a funnel: &lt;code&gt;FilterEvaluated&lt;/code&gt; → &lt;code&gt;FilterMatched&lt;/code&gt; → &lt;code&gt;EventsDelivered&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Pricing: cheaper, until your events get big
&lt;/h2&gt;

&lt;p&gt;AWS's worked example (1,000 events/s at 2 KB, two subscribers, JSONata plus schema validation, 7-day retention, 1% failures) comes to &lt;strong&gt;$2,317/month&lt;/strong&gt;: $922 ingress, $518 egress, $16 retry egress, $778 evaluation, $83 storage. The math checks out.&lt;/p&gt;

&lt;p&gt;What the announcement doesn't show is the comparison with classic, which charges $1 per million events in 64 KB chunks, delivery within an account free and $1 per million for each cross-account delivery. Same volume (2.592 billion events/month, two subscribers, ingress and egress only):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Event size&lt;/th&gt;
&lt;th&gt;New bus&lt;/th&gt;
&lt;th&gt;Classic (same account / 2 cross-account deliveries)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;1 KB&lt;/td&gt;
&lt;td&gt;~$726&lt;/td&gt;
&lt;td&gt;$2,592 / $7,776&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2 KB&lt;/td&gt;
&lt;td&gt;~$1,440&lt;/td&gt;
&lt;td&gt;$2,592 / $7,776&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;10 KB&lt;/td&gt;
&lt;td&gt;~$6,002&lt;/td&gt;
&lt;td&gt;$2,592 / $7,776&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;60 KB&lt;/td&gt;
&lt;td&gt;~$34,514&lt;/td&gt;
&lt;td&gt;$2,592 / $7,776&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;My own math from published prices; excludes retention, retries, evaluation, and classic archives.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Small events crossing accounts get much cheaper. Large events can get several times more expensive, because classic charges the same for 1 KB and 64 KB. Against same-account classic with two subscribers, break-even sits around 4 KB per event. So slim your events (claim-check pattern), consider Avro or Protobuf via &lt;code&gt;PutRawEvents&lt;/code&gt;, be deliberate about paid evaluations, and size retention to what you'll actually replay.&lt;/p&gt;

&lt;h2&gt;
  
  
  Availability
&lt;/h2&gt;

&lt;p&gt;Launch Regions: N. Virginia, Ohio, Oregon, Ireland, Frankfurt, Stockholm, Spain, Hong Kong, Malaysia, Mumbai, Singapore, Sydney, Thailand and Tokyo. No São Paulo. Because subscribers must be in the bus's Region, South American workloads would need to bridge from a classic bus to a new bus in another Region. That's possible, but it adds latency, inter-Region transfer costs and data residency questions, especially now that the bus stores events.&lt;/p&gt;

&lt;h2&gt;
  
  
  My take
&lt;/h2&gt;

&lt;p&gt;This is the biggest change to EventBridge since it launched, and ownership matters more than ordering here. Letting each team create and pay for subscribers on a bus it doesn't own is what makes a central bus something the platform team operates instead of something it serves by ticket. Per-subscriber replay from built-in retention is the other big win. Before migrating, run your real event sizes through the pricing, understand the per-group blocking on ordered subscribers, and check Region coverage. Start with a new domain, or one that's hurting for ordering or replay, and leave the rest on classic, which isn't going anywhere.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published in Spanish on &lt;a href="https://blog.cloudacademy.ar/eventbridge-nuevo-custom-event-bus-orden-replay-organizacion/" rel="noopener noreferrer"&gt;CloudAcademy.ar&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/blogs/aws/introducing-enhanced-custom-event-buses-in-amazon-eventbridge-for-enterprise-scale-event-driven-applications/" rel="noopener noreferrer"&gt;AWS News Blog announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/09/eventbridge-relaunches-custom-event-buses/" rel="noopener noreferrer"&gt;What's New post&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/eventbridge/latest/userguide/eb-custom-bus-what-is.html" rel="noopener noreferrer"&gt;Custom Event Bus documentation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://docs.aws.amazon.com/eventbridge/latest/userguide/eb-custom-bus-subscribers.html" rel="noopener noreferrer"&gt;Subscribers&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/eventbridge/latest/userguide/eb-custom-bus-retry.html" rel="noopener noreferrer"&gt;Retries and ordering&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/eventbridge/latest/userguide/eb-custom-bus-replay.html" rel="noopener noreferrer"&gt;Replay&lt;/a&gt; · &lt;a href="https://docs.aws.amazon.com/eventbridge/latest/userguide/eb-custom-bus-sharing.html" rel="noopener noreferrer"&gt;Sharing&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/eventbridge/pricing/" rel="noopener noreferrer"&gt;EventBridge pricing&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>serverless</category>
      <category>eventdriven</category>
      <category>architecture</category>
    </item>
    <item>
      <title>Amazon GuardDuty AI-powered investigations: a SOC copilot, not an autopilot</title>
      <dc:creator>Matias Martinez</dc:creator>
      <pubDate>Mon, 29 Jun 2026 22:51:46 +0000</pubDate>
      <link>https://dev.to/matias_martinez_185d9e0ee/amazon-guardduty-ai-powered-investigations-a-soc-copilot-not-an-autopilot-eah</link>
      <guid>https://dev.to/matias_martinez_185d9e0ee/amazon-guardduty-ai-powered-investigations-a-soc-copilot-not-an-autopilot-eah</guid>
      <description>&lt;h1&gt;
  
  
  Amazon GuardDuty AI-powered investigations: a useful SOC copilot, not an autopilot
&lt;/h1&gt;

&lt;p&gt;AWS just added a capability that many security teams have been trying to build around GuardDuty for years: a faster way to move from “there is a finding” to “this is probably what happened, here is the evidence, and these are the next actions.” The new feature is called &lt;strong&gt;GuardDuty Investigation&lt;/strong&gt;, and it is now available in preview as &lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/06/amazon-guardduty/" rel="noopener noreferrer"&gt;AI-powered investigations for Amazon GuardDuty&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;That wording matters. GuardDuty has already used machine learning and threat intelligence to detect suspicious behavior across AWS accounts, EC2, EKS, ECS, Lambda, S3, RDS, and other surfaces. This launch is not only about detecting another signal. It is about reducing the investigation tax that comes after the alert.&lt;/p&gt;

&lt;p&gt;If you have operated cloud security at scale, you know the pattern: a finding appears, the analyst checks CloudTrail, IAM context, affected resources, related activity, known indicators, suppression rules, account ownership, and whether the finding is a real incident or just a noisy but expected behavior. That work is valuable, but it is also repetitive. It is exactly where a well-scoped AI assistant can help.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2s4pry5davvj8o9c8q0k.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2s4pry5davvj8o9c8q0k.png" alt="Amazon GuardDuty overview" width="800" height="335"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  🔵 What changed
&lt;/h2&gt;

&lt;p&gt;According to the &lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/06/amazon-guardduty/" rel="noopener noreferrer"&gt;AWS announcement&lt;/a&gt;, AI-powered investigations automatically analyze GuardDuty findings and AWS accounts to help distinguish true threats from benign findings. The investigation looks at finding context, related activity from the last 90 days, affected resources, and threat indicators. AWS says it uses knowledge graphs and threat intelligence to produce the analysis in minutes.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://docs.aws.amazon.com/guardduty/latest/ug/guardduty-investigation.html" rel="noopener noreferrer"&gt;GuardDuty Investigation documentation&lt;/a&gt; gives more detail on the output. Each investigation can produce:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a risk level: Info, Low, Medium, High, or Critical;&lt;/li&gt;
&lt;li&gt;a confidence level: Unknown, Low, Medium, or High;&lt;/li&gt;
&lt;li&gt;a summary with key observations;&lt;/li&gt;
&lt;li&gt;investigation details with supporting context;&lt;/li&gt;
&lt;li&gt;recommended actions, including CLI commands;&lt;/li&gt;
&lt;li&gt;MITRE ATT&amp;amp;CK technique classification, using the common security knowledge base for attacker tactics and techniques.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That is a meaningful step beyond a plain alert. It gives the responder a starting hypothesis and evidence trail. The important part is not that the model “knows security.” The useful part is that GuardDuty already has the AWS-native context and can package that context into a structured investigation.&lt;/p&gt;

&lt;h2&gt;
  
  
  🔶 Three analysis modes
&lt;/h2&gt;

&lt;p&gt;The preview supports three investigation scopes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Finding analysis&lt;/strong&gt; focuses on one GuardDuty finding. You provide a 32-character finding ID, and GuardDuty analyzes that specific signal. During preview, AWS says this supports all Extended Threat Detection findings and selected findings from foundational, S3, and Runtime plans.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Account analysis&lt;/strong&gt; looks at the threat posture of one AWS account. This is useful when an account has multiple signals or when the first question is not “is this one finding real?” but “is this account behaving like it is under attack?”&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Organization analysis&lt;/strong&gt; lets an administrator analyze organization-wide posture, with a preview limit of up to 100 accounts. This is the most interesting mode for larger AWS environments because cloud incidents rarely respect account boundaries. A compromised principal, suspicious API activity, or lateral movement pattern may only make sense when viewed across accounts.&lt;/p&gt;

&lt;p&gt;That said, the preview limits are real. The documentation states that, during preview, you can initiate up to 10 investigations per account per day, with a total limit of 100 investigations per account. The trigger prompt for API or CLI usage can be up to 2,048 characters. Failed investigations do not count against those quotas.&lt;/p&gt;

&lt;h2&gt;
  
  
  ⚙️ How to enable it
&lt;/h2&gt;

&lt;p&gt;You need an active GuardDuty detector in the Region where you want to create investigations, and the investigation feature must be enabled for that detector. In the console, AWS exposes this under &lt;strong&gt;Settings → AI powered investigations - Preview&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;From the CLI, the docs show the feature name as &lt;code&gt;AI_ANALYST&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws guardduty update-detector &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--detector-id&lt;/span&gt; 2cb3d4e5f6a7b8c9d0e1f2a3b4c5d6e7 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--features&lt;/span&gt; &lt;span class="s1"&gt;'[{"Name":"AI_ANALYST","Status":"ENABLED"}]'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Your IAM identity also needs the investigation permissions:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"Version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"2012-10-17"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"Statement"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Effect"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Allow"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Action"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="s2"&gt;"guardduty:CreateInvestigation"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="s2"&gt;"guardduty:GetInvestigation"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="s2"&gt;"guardduty:ListInvestigations"&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"Resource"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"arn:aws:guardduty:us-west-2:123456789012:detector/2cb3d4e5f6a7b8c9d0e1f2a3b4c5d6e7"&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A basic investigation can be started with &lt;code&gt;create-investigation&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws guardduty create-investigation &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--detector-id&lt;/span&gt; 2cb3d4e5f6a7b8c9d0e1f2a3b4c5d6e7 &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--trigger-prompt&lt;/span&gt; &lt;span class="s2"&gt;"Investigate finding 1ab2c3d4e5f6a7b8c9d0e1f2a3b4c5d6 in account 123456789012"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The API is asynchronous. You get an &lt;code&gt;InvestigationId&lt;/code&gt;, then retrieve the result with &lt;code&gt;get-investigation&lt;/code&gt; once it completes.&lt;/p&gt;

&lt;h2&gt;
  
  
  🛡️ Where this helps in practice
&lt;/h2&gt;

&lt;p&gt;The obvious use case is SOC triage. Instead of every finding starting as a blank page, the analyst receives a structured summary with risk, confidence, evidence, and recommended commands. That can cut the first 20–40 minutes of manual digging, especially for teams that already receive more alerts than they can deeply review.&lt;/p&gt;

&lt;p&gt;The second use case is consistency. Different analysts may investigate the same GuardDuty finding differently. A standardized AI-generated package can help normalize the first response: what was checked, why the risk was assigned, which known attack technique may apply, and which resources need attention.&lt;/p&gt;

&lt;p&gt;The third use case is managed security. If you operate AWS Organizations for many teams, account analysis and organization analysis are more interesting than single-finding analysis. They can help a central security team answer “which accounts deserve attention first?” instead of treating every alert as equal.&lt;/p&gt;

&lt;h2&gt;
  
  
  ⚠️ What to be careful about
&lt;/h2&gt;

&lt;p&gt;The preview documentation is explicit: investigation recommendations may contain errors or incomplete assessments, and human review is recommended. That is the right mental model. This feature should not auto-remediate production resources without a review gate.&lt;/p&gt;

&lt;p&gt;There is also a data residency detail worth reading carefully. GuardDuty Investigation uses Cross-Region Inference Service. AWS says your data remains stored only in the Region where the investigation request originates, but investigation data and summary results may be processed outside that Region within the supported geography. For many organizations that is acceptable; for regulated environments, it needs review.&lt;/p&gt;

&lt;p&gt;Availability is also limited to 10 commercial Regions in preview: US East (N. Virginia), US East (Ohio), US West (Oregon), Canada (Central), Europe (Frankfurt), Europe (Ireland), Europe (London), Europe (Paris), Europe (Stockholm), and Asia Pacific (Tokyo).&lt;/p&gt;

&lt;h2&gt;
  
  
  🔍 What this means for security teams
&lt;/h2&gt;

&lt;p&gt;This is the right direction for cloud security tooling. The value is not “AI replaces the analyst.” The value is that AWS has the telemetry, the resource graph, the threat intelligence, and the service context. If GuardDuty can assemble that into a defensible investigation package, analysts can spend less time collecting facts and more time making decisions.&lt;/p&gt;

&lt;p&gt;I would not wire this directly to remediation on day one. I would start by enabling it in a controlled set of accounts, comparing its summaries against human investigations, measuring time saved, and building a review workflow around the recommendations. If the results are good, the next step is not full autopilot. It is faster triage, better evidence, and cleaner escalation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/about-aws/whats-new/2026/06/amazon-guardduty/" rel="noopener noreferrer"&gt;AWS What's New: Amazon GuardDuty AI-powered investigations accelerate threat response&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://docs.aws.amazon.com/guardduty/latest/ug/guardduty-investigation.html" rel="noopener noreferrer"&gt;Amazon GuardDuty User Guide: GuardDuty Investigation&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://aws.amazon.com/guardduty/" rel="noopener noreferrer"&gt;Amazon GuardDuty product page&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://iamtrail.com/guardduty" rel="noopener noreferrer"&gt;IAMTrail GuardDuty announcements archive&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.secureresearch.com/tech-articles/2026-06-23-amazon-guardduty-ai-powered-investigations-acceler-9fab34a357f5.html" rel="noopener noreferrer"&gt;SecureResearch summary of the AWS announcement&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>aws</category>
      <category>security</category>
      <category>cloud</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
