๐ November 5, 2026 โข AWS Machine Learning Blog
Amazon has moved Web Answer to general availability on Bedrock AgentCore, according to the AWS machine learning blog post published on November 5. The service packages the full search-retrieve-synthesize-cite loop behind a single managed tool call, so an application no longer has to orchestrate the individual steps itself.
The mechanics are straightforward. An agent sends one natural-language question as one MCP tool call to an AgentCore Gateway and receives a synthesized answer in which each claim is mapped to a source URL. Under the hood the system runs agentic search autonomously. It plans its own queries, runs several of them in parallel, reads key pages to verify facts, and searches again when the evidence is insufficient. The model is given today's date as part of the request context, which keeps results current.
For engineers wiring agents to internal or external knowledge, the practical question is usually where the retrieval loop should live. Doing it in-house means owning query planning, parallel fetches, page reading and citation tracking as application code. Handing it to Web Answer means those steps run inside the managed service and arrive as a single response, trading control for a shorter integration path.
Pricing is per call. Web Answer costs $10 per 1,000 tool calls, which makes cost estimation for an agent workload a simple multiplication of call volume. Availability starts in one region, US East (N. Virginia), with more regions to come, so latency-sensitive deployments outside that region should plan for cross-region calls until expansion lands. The Gateway authenticates calls with IAM or JWT, so the tool fits the standard access-control patterns already used with AgentCore.
AWS also draws a line between the two retrieval surfaces. According to the post, Web Search returns ranked raw results for your own model to reason over, while Web Answer runs the retrieval loop and the synthesis for you and hands back a cited answer. The choice is between keeping reasoning in your model with raw material, or delegating the research loop to the managed service and consuming its output directly.
IMPACT
A retrieval pipeline that used to be several steps of planning, fetching and citation bookkeeping collapses into one MCP tool call with claims mapped to URLs. At $10 per 1,000 calls it is easy to model cost, but check region coverage, since GA starts only in US East (N. Virginia).
Sources: AWS Machine Learning Blog
Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky
Top comments (0)