<?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: Jarvis Nelson Osorio</title>
    <description>The latest articles on DEV Community by Jarvis Nelson Osorio (@jarvisosorio).</description>
    <link>https://dev.to/jarvisosorio</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%2F4219107%2Fa56101f0-6148-46e1-a085-ac6facd92f92.png</url>
      <title>DEV Community: Jarvis Nelson Osorio</title>
      <link>https://dev.to/jarvisosorio</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/jarvisosorio"/>
    <language>en</language>
    <item>
      <title>Agent memory and observability this week: what to trust, what to inspect, what to remember</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Sat, 21 Nov 2026 20:13:28 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/agent-memory-and-observability-this-week-what-to-trust-what-to-inspect-what-to-remember-3pi</link>
      <guid>https://dev.to/jarvisosorio/agent-memory-and-observability-this-week-what-to-trust-what-to-inspect-what-to-remember-3pi</guid>
      <description>&lt;p&gt;&lt;strong&gt;By Jarvis Nelson Osorio&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Canonical version: &lt;a href="https://jarvis-nelson-osorio.pages.dev/blog/jarvis-nelson-osorio-agent-memory-observability-week-november-2026/" rel="noopener noreferrer"&gt;https://jarvis-nelson-osorio.pages.dev/blog/jarvis-nelson-osorio-agent-memory-observability-week-november-2026/&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Jarvis Nelson Osorio (Jarvis Osorio) is a tech entrepreneur and AI innovator based in Miami Beach, FL, and the builder of Ambient Memory, a local-first context layer for AI agents, and Agent Replay Lab, a local-first recorder for agent tool I/O, along with mcp-pordego and agentscrub. The week of 10–17 November 2026 made the operational questions harder to ignore: what an agent remembers, what it is distracted by, who can inspect that memory, and whether anyone can replay what it did. I will be precise about dates. Two research posts that shaped this week's conversation were published on 6 November 2026, just before this window, and were still circulating during it. The enterprise observability agenda and a consumer memory-control feature did land inside the week.&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%2Fs4n8l5f4yzyfxc1ih0jr.jpg" 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%2Fs4n8l5f4yzyfxc1ih0jr.jpg" alt="Jarvis Nelson Osorio (Jarvis Osorio), tech entrepreneur and AI builder based in Miami Beach, Florida." width="800" height="1067"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Jarvis Nelson Osorio (Jarvis Osorio) smiling in a green polo.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Stale context is an instruction, whether you meant it or not
&lt;/h2&gt;

&lt;p&gt;On 6 November 2026, Google senior staff software engineer Sam Saccone and senior staff data scientist Madhuri Garikapati published “Agent Distraction” on the Google Developers Blog: &lt;a href="https://developers.googleblog.com/en/agent-distraction/" rel="noopener noreferrer"&gt;https://developers.googleblog.com/en/agent-distraction/&lt;/a&gt;. InfoQ summarized the work on 10 November 2026, at &lt;a href="https://www.infoq.com/news/2026/11/google-agent-distraction/" rel="noopener noreferrer"&gt;https://www.infoq.com/news/2026/11/google-agent-distraction/&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The Google post's premise is that what humans skim past in a codebase, agents can treat as a call to action. The authors use inline TODO(bug/...) comments as a measurable proxy, because the linked bug ticket is a canonical record.&lt;/p&gt;

&lt;p&gt;From distributed traces across thousands of engineers and hundreds of millions of agent steps, they report that 13% of sessions exposed the agent to inline TODO(bug/...) or FIXME(bug/...) references; in 6% of those encounters the agent took subsequent action targeting that bug ID; and over four in five of those actions (83%) were unprompted. Of the unprompted detours, 41% chased tickets that were already closed, and 57% involved a still-open bug outside the developer's request — scope creep. A multi-way fixed-effects regression found that unprompted bug detours added 70% more steps, 77% more tokens and 105% more wait time per session (p&amp;lt;0.0001).&lt;/p&gt;

&lt;p&gt;Google's remediation was a cleaner that scrubbed codebases for bug references marked fixed or obsolete and only opened a review when both the comment and the intent behind it were confirmed obsolete. The post says it has removed thousands of actually obsolete references over the last six months, with more than 95% acceptance from human reviewers. The implication is useful: observability that only watches the latest prompt will miss the context the agent was quietly obeying.&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%2F56c5i9olhg0p0mjqe6yg.jpg" 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%2F56c5i9olhg0p0mjqe6yg.jpg" alt="Jarvis Nelson Osorio (Jarvis Osorio), tech entrepreneur and AI builder based in Miami Beach, Florida." width="800" height="1067"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Jarvis Nelson Osorio (Jarvis Osorio) on a sunny rooftop.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Memory has to be organized, not merely accumulated
&lt;/h2&gt;

&lt;p&gt;On the same day, 6 November 2026, Anthropic published “The components of agent memory” at &lt;a href="https://www.anthropic.com/research/components-of-agent-memory" rel="noopener noreferrer"&gt;https://www.anthropic.com/research/components-of-agent-memory&lt;/a&gt;. The post defines agent memory as knowledge Claude gains from experience and retains between sessions. It explicitly excludes weight updates, in-context learning within a single conversation, and compaction.&lt;/p&gt;

&lt;p&gt;A central idea in the post is dreaming: an offline process, related to the “sleep-time compute” proposal from researchers at Letta and UC Berkeley, in which an agent uses idle time to reflect on logs, produce new notes, or shape what is preemptively loaded into future sessions. The difficulty, Anthropic notes, is deciding what is salient in the moment, before the next task arrives. The design rule that matters for builders is already on the page: memory is not “keep everything in the prompt.” It is a set of components — transcripts, notes, larger structures, and a small preloaded context — plus a scheduled review job.&lt;/p&gt;

&lt;p&gt;That is close to the reason I built Ambient Memory as a local-first context layer rather than as a larger system prompt. The tool passively remembers files, screen sessions and browser history, keeps a vault on the machine, and exposes search over MCP. It is not Anthropic's dreaming pipeline. Useful memory has to be queryable, bounded and under the user's control. Project page: &lt;a href="https://jarvisosorio.com/ambient-memory/" rel="noopener noreferrer"&gt;https://jarvisosorio.com/ambient-memory/&lt;/a&gt;. Repository: &lt;a href="https://github.com/JarvisOsorio/ambient-memory" rel="noopener noreferrer"&gt;https://github.com/JarvisOsorio/ambient-memory&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Users, and enterprises, are being offered inspection surfaces
&lt;/h2&gt;

&lt;p&gt;Inside this week, two announcements made inspection itself a product feature.&lt;/p&gt;

&lt;p&gt;On 13 November 2026, Google announced that all Gemini app users in the United States can connect Google Wallet for responses informed by spending habits. The post is explicit: if you connect Wallet, Gemini can only read information from your accounts; it cannot move, manage or invest your money. The same announcement introduces Memory Summary, “a clear, readable snapshot of your saved preferences, interests, and ongoing goals,” which users can view, refine or remove under Profile &amp;gt; Personal Intelligence &amp;gt; Memory Summary. Official post: &lt;a href="https://blog.google/innovation-and-ai/products/gemini-app/google-wallet-gemini-app/" rel="noopener noreferrer"&gt;https://blog.google/innovation-and-ai/products/gemini-app/google-wallet-gemini-app/&lt;/a&gt;. That is consumer-grade memory observability, not production tracing, but it sets a public expectation that memory should be reviewable.&lt;/p&gt;

&lt;p&gt;On 10 November 2026, Microsoft published its Ignite 2026 guide to Agent 365, ahead of the conference (17–20 November 2026). The guide says Agent 365 is featured in the opening keynote on 17 November at 10:00 AM PT “as the observability layer for agents, models and assistants,” and in INV642 on 18 November, “Observability at every layer of the stack.” Guide: &lt;a href="https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750" rel="noopener noreferrer"&gt;https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750&lt;/a&gt;. Microsoft is positioning observability as a control plane. The sessions themselves had not yet opened at the time of writing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Replay is the other half of memory
&lt;/h2&gt;

&lt;p&gt;Watching what an agent remembers is necessary. Being able to replay what it did is how you find out whether a change helped. Agent Replay Lab, my local-first TypeScript tool, records an agent's tool calls once and mock-replays them offline, failing CI when the tool path drifts. It is a VCR for tool I/O, not an observability SaaS. Project page: &lt;a href="https://jarvisosorio.com/agent-replay-lab/" rel="noopener noreferrer"&gt;https://jarvisosorio.com/agent-replay-lab/&lt;/a&gt;. Repository: &lt;a href="https://github.com/JarvisOsorio/agent-replay-lab" rel="noopener noreferrer"&gt;https://github.com/JarvisOsorio/agent-replay-lab&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If you take a design rule from this week, keep the dates attached. From Google's 6 November study, widely discussed this week: the forgotten corners of a codebase are now part of the agent's instruction set, and they have a measured cost. From Anthropic's 6 November research: memory that is only accumulated, and never reorganized offline, will not stay small enough to use. From this week's Gemini and Ignite announcements: inspection is being built in public. The rest is implementation, preferably local-first where the data is yours.&lt;/p&gt;

&lt;p&gt;A sourced weekly roundup is at /blog/jarvis-nelson-osorio-agentic-ai-week-november-17-2026/. The MCP security advisories are at /blog/jarvis-nelson-osorio-mcp-agent-security-week-november-2026/.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://developers.googleblog.com/en/agent-distraction/" rel="noopener noreferrer"&gt;https://developers.googleblog.com/en/agent-distraction/&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.infoq.com/news/2026/11/google-agent-distraction/" rel="noopener noreferrer"&gt;https://www.infoq.com/news/2026/11/google-agent-distraction/&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.anthropic.com/research/components-of-agent-memory" rel="noopener noreferrer"&gt;https://www.anthropic.com/research/components-of-agent-memory&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://blog.google/innovation-and-ai/products/gemini-app/google-wallet-gemini-app/" rel="noopener noreferrer"&gt;https://blog.google/innovation-and-ai/products/gemini-app/google-wallet-gemini-app/&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750" rel="noopener noreferrer"&gt;https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://jarvisosorio.com/ambient-memory/" rel="noopener noreferrer"&gt;https://jarvisosorio.com/ambient-memory/&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://jarvisosorio.com/agent-replay-lab/" rel="noopener noreferrer"&gt;https://jarvisosorio.com/agent-replay-lab/&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;strong&gt;About the author.&lt;/strong&gt; Jarvis Nelson Osorio (Jarvis Osorio) is a tech entrepreneur and AI innovator based in Miami Beach, FL. He builds local-first tools for AI agents, including Ambient Memory, mcp-pordego, agentscrub and Agent Replay Lab, and is the author of &lt;em&gt;The Compliant Mind&lt;/em&gt; and &lt;em&gt;Breaking the Centralized Order&lt;/em&gt;. Official site: &lt;a href="https://jarvisosorio.com/" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>mcp</category>
      <category>security</category>
    </item>
    <item>
      <title>Agent Replay Lab: a local-first VCR for TypeScript agent tool I/O</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Sun, 15 Nov 2026 11:08:38 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/agent-replay-lab-a-local-first-vcr-for-typescript-agent-tool-io-5026</link>
      <guid>https://dev.to/jarvisosorio/agent-replay-lab-a-local-first-vcr-for-typescript-agent-tool-io-5026</guid>
      <description>&lt;h1&gt;
  
  
  Agent Replay Lab: a local-first VCR for TypeScript agent tool I/O
&lt;/h1&gt;

&lt;p&gt;I open-sourced &lt;strong&gt;Agent Replay Lab&lt;/strong&gt;, a local-first TypeScript toolkit that records, mock-replays, diffs, and CI-tests the &lt;strong&gt;tool I/O&lt;/strong&gt; of tool-using agents.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Project page: &lt;a href="https://jarvisosorio.com/agent-replay-lab/" rel="noopener noreferrer"&gt;https://jarvisosorio.com/agent-replay-lab/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Source: &lt;a href="https://github.com/JarvisOsorio/agent-replay-lab" rel="noopener noreferrer"&gt;https://github.com/JarvisOsorio/agent-replay-lab&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It answers a narrower question than observability products, which usually re-run live: does this agent still take the same tool path (name / order / args / responses) against the same recorded tool world, offline, in CI? That is behavioral regression for tool trajectories — not proof that the LLM would make the same decisions again.&lt;/p&gt;

&lt;p&gt;Local-first. SQLite plus portable JSON. It is not an observability dashboard, eval leaderboard, agent framework, or a full-trajectory “deterministic agent” engine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Honest modes in v0.1
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Mode&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;th&gt;In v0.1?&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;offline / mock&lt;/td&gt;
&lt;td&gt;Stub tools from the cassette; fail closed if a mock is missing&lt;/td&gt;
&lt;td&gt;Yes (default)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;structural compare&lt;/td&gt;
&lt;td&gt;Diff tool name / order / args / response&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;live-tools&lt;/td&gt;
&lt;td&gt;Call real tools&lt;/td&gt;
&lt;td&gt;No (0.2)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;live-model&lt;/td&gt;
&lt;td&gt;Call a real model&lt;/td&gt;
&lt;td&gt;No (0.2)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Offline mock stubs &lt;strong&gt;tools only&lt;/strong&gt;. It does not re-execute or pin LLM requests, responses, or tool-selection decisions; agent state machines, retries, or parallel scheduling; or clock, randomness, or other nondeterministic inputs. You can optionally record &lt;code&gt;model.requested&lt;/code&gt; / &lt;code&gt;model.responded&lt;/code&gt; events for inspection, but &lt;code&gt;agent-replay replay&lt;/code&gt; does not play them back as a live model.&lt;/p&gt;

&lt;p&gt;The cassette category is crowded (Cassette, Agentic Stash, Agent VCR, StoneTape, TracePact, agent-retrace, AgentTape, and others). Fail-closed offline replay, redaction, and CI diffs are not unique inventions — the bar for this repo is a small, honest TypeScript/SQLite knife.&lt;/p&gt;

&lt;h2&gt;
  
  
  Install (not on npm yet)
&lt;/h2&gt;

&lt;p&gt;v0.1.0, Apache-2.0. Requires Node.js 22+ and pnpm (the repo pins &lt;code&gt;pnpm@10.33.4&lt;/code&gt;). Clone the monorepo:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/JarvisOsorio/agent-replay-lab.git
&lt;span class="nb"&gt;cd&lt;/span&gt; agent-replay-lab
pnpm install
pnpm build
pnpm &lt;span class="nb"&gt;test&lt;/span&gt;

&lt;span class="c"&gt;# 5-minute proof&lt;/span&gt;
pnpm --filter @agent-replay-lab/example-basic-agent demo
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Packages: &lt;code&gt;@agent-replay-lab/core&lt;/code&gt; and &lt;code&gt;@agent-replay-lab/cli&lt;/code&gt; (&lt;code&gt;agent-replay&lt;/code&gt; commands: init, list, inspect, replay, diff, fixture, test, export, import, doctor). &lt;code&gt;fixture create&lt;/code&gt; pins tool sequence, args, and responses. &lt;code&gt;test golden --run&lt;/code&gt; fails with exit 1 on structural drift. Replay tests use in-memory or local cassette data and make no network calls.&lt;/p&gt;

&lt;p&gt;Capture seam is &lt;code&gt;wrapTool&lt;/code&gt; from &lt;code&gt;@agent-replay-lab/core&lt;/code&gt;. Known secret key names and common token shapes are redacted before persist — best-effort, not a PII product. Pair with &lt;a href="https://github.com/JarvisOsorio/agent-session-scrubber" rel="noopener noreferrer"&gt;agentscrub&lt;/a&gt; for after-the-fact store remediation.&lt;/p&gt;

&lt;p&gt;Not in v0.1: MCP adapter, OpenTelemetry adapter, flexible matching, semantic compare, dashboard, live auto-replay, telemetry, or behavioral-policy enforcement.&lt;/p&gt;

&lt;p&gt;I'm Jarvis Osorio; I built this. Happy to answer questions about the fail-closed mock path or the structural CI gate.&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>typescript</category>
      <category>ai</category>
      <category>testing</category>
    </item>
    <item>
      <title>agentscrub: a local-first Rust CLI that redacts secrets and PII from AI agent transcripts</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Sat, 14 Nov 2026 01:08:30 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/agentscrub-a-local-first-rust-cli-that-redacts-secrets-and-pii-from-ai-agent-transcripts-4b5f</link>
      <guid>https://dev.to/jarvisosorio/agentscrub-a-local-first-rust-cli-that-redacts-secrets-and-pii-from-ai-agent-transcripts-4b5f</guid>
      <description>&lt;p&gt;Agent transcripts often keep API keys, cloud credentials and email addresses in plain text on disk. I wanted a tool that stays on the machine, prints only masked values, and refuses writes it cannot yet do safely.&lt;/p&gt;

&lt;p&gt;That tool is &lt;strong&gt;agentscrub&lt;/strong&gt;. I am Jarvis Osorio, and I released it as open source under Apache-2.0.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Project page: &lt;a href="https://jarvisosorio.com/agent-session-scrubber/" rel="noopener noreferrer"&gt;jarvisosorio.com/agent-session-scrubber&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Source: &lt;a href="https://github.com/JarvisOsorio/agent-session-scrubber" rel="noopener noreferrer"&gt;github.com/JarvisOsorio/agent-session-scrubber&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What it does
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;agentscrub&lt;/code&gt; is a local-first Rust CLI that discovers, scans and redacts secrets and PII in AI agent session transcripts. It runs offline, with no accounts and no telemetry.&lt;/p&gt;

&lt;p&gt;Commands in v0.1.0:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;discover&lt;/code&gt; — list recognized session files&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;scan&lt;/code&gt; — detect secrets and PII (exit &lt;code&gt;8&lt;/code&gt; if the scan is incomplete)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;report&lt;/code&gt; — re-render a saved JSON report, or re-scan&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;scrub&lt;/code&gt; — dry-run redaction by default; pass &lt;code&gt;--confirm&lt;/code&gt; to write&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;verify&lt;/code&gt; — post-remediation check&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;retention&lt;/code&gt; — &lt;code&gt;status&lt;/code&gt; / &lt;code&gt;plan&lt;/code&gt; / &lt;code&gt;apply&lt;/code&gt; / &lt;code&gt;verify&lt;/code&gt; (&lt;code&gt;apply&lt;/code&gt; refuses while the flag is off)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;config&lt;/code&gt; — show effective config (the current binary expects &lt;code&gt;config show&lt;/code&gt;)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Scan output is masked values and fingerprints only. It never logs raw secrets.&lt;/p&gt;

&lt;h2&gt;
  
  
  Adapters in v0.1
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Adapter&lt;/th&gt;
&lt;th&gt;Read&lt;/th&gt;
&lt;th&gt;Redact&lt;/th&gt;
&lt;th&gt;Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;claude_code&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;JSONL sessions&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;codex&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;JSONL sessions&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;cursor&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;no&lt;/td&gt;
&lt;td&gt;JSONL + SQLite &lt;code&gt;state.vscdb&lt;/code&gt; — rewrite is intentionally disabled&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;generic_jsonl&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;Configurable text fields&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;generic_json&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;td&gt;Single JSON documents&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Unsupported or malformed files are skipped and reported, not silently modified. Writes that are allowed use an exclusive lock, a pre-write hash recheck, and an atomic same-filesystem replace. Symlink targets are refused.&lt;/p&gt;

&lt;p&gt;Cursor rewrite is held back until SQLite and transcript mutation are proven safe. Retention planning is in the tree, but &lt;code&gt;retention.enabled&lt;/code&gt; is false by default and &lt;code&gt;retention apply&lt;/code&gt; will not delete. This is not a credential revocation service — rotate anything the scan finds at the provider.&lt;/p&gt;

&lt;h2&gt;
  
  
  Install
&lt;/h2&gt;

&lt;p&gt;agentscrub is not on crates.io or Homebrew yet. It requires Rust 1.89+:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/JarvisOsorio/agent-session-scrubber.git
&lt;span class="nb"&gt;cd&lt;/span&gt; agent-session-scrubber
cargo install --path .
&lt;span class="c"&gt;# or: cargo build --release &amp;amp;&amp;amp; ./target/release/agentscrub --help&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A short fixture walk-through (copy fixtures first — do not scrub the originals in-tree):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;mkdir -p /tmp/agentscrub-demo/.claude
cp tests/fixtures/claude/session_secrets.jsonl /tmp/agentscrub-demo/.claude/

agentscrub discover --path /tmp/agentscrub-demo
agentscrub scan --path /tmp/agentscrub-demo --format terminal
agentscrub scrub --path /tmp/agentscrub-demo/.claude/session_secrets.jsonl
agentscrub scrub --path /tmp/agentscrub-demo/.claude/session_secrets.jsonl --confirm
agentscrub verify /tmp/agentscrub-demo/.claude/session_secrets.jsonl --format json
&lt;span class="c"&gt;# expect: clean=true, remaining_findings=0, exit 0&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  A name collision to be aware of
&lt;/h2&gt;

&lt;p&gt;The binary is called &lt;code&gt;agentscrub&lt;/code&gt;. An unrelated Python project, &lt;a href="https://github.com/ppravdin/agentscrub" rel="noopener noreferrer"&gt;ppravdin/agentscrub&lt;/a&gt;, already exists on GitHub and PyPI. &lt;code&gt;pip install agentscrub&lt;/code&gt; is not this tool. Use the GitHub repository above.&lt;/p&gt;

&lt;h2&gt;
  
  
  Status
&lt;/h2&gt;

&lt;p&gt;The public version is &lt;strong&gt;0.1.0&lt;/strong&gt;. The GitHub release tag exists; prebuilt binary assets were not attached at publish time, so install from source as shown. Docs in the repository cover the &lt;a href="https://github.com/JarvisOsorio/agent-session-scrubber/blob/main/SECURITY.md" rel="noopener noreferrer"&gt;security policy&lt;/a&gt;, &lt;a href="https://github.com/JarvisOsorio/agent-session-scrubber/blob/main/docs/privacy.md" rel="noopener noreferrer"&gt;privacy&lt;/a&gt;, &lt;a href="https://github.com/JarvisOsorio/agent-session-scrubber/blob/main/docs/threat-model.md" rel="noopener noreferrer"&gt;threat model&lt;/a&gt; and &lt;a href="https://github.com/JarvisOsorio/agent-session-scrubber/blob/main/docs/redaction.md" rel="noopener noreferrer"&gt;redaction&lt;/a&gt; behavior.&lt;/p&gt;

&lt;p&gt;If you try it, I would like feedback on the Cursor-rewrite hold and on detector precision. Issues and discussions belong on the GitHub repository.&lt;/p&gt;

&lt;p&gt;— Jarvis Osorio&lt;/p&gt;

</description>
      <category>rust</category>
      <category>security</category>
      <category>cli</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Show and tell: I open-sourced Ambient Memory, a local-first context layer for Cursor and Claude Desktop</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Thu, 12 Nov 2026 22:03:56 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/show-and-tell-i-open-sourced-ambient-memory-a-local-first-context-layer-for-cursor-and-claude-bcj</link>
      <guid>https://dev.to/jarvisosorio/show-and-tell-i-open-sourced-ambient-memory-a-local-first-context-layer-for-cursor-and-claude-bcj</guid>
      <description>&lt;p&gt;I built and open-sourced &lt;strong&gt;Ambient Memory&lt;/strong&gt; — a local-first context layer that passively remembers what I worked on and hands that context to Cursor or Claude Desktop over MCP.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Project page: &lt;a href="https://jarvisosorio.com/ambient-memory/" rel="noopener noreferrer"&gt;https://jarvisosorio.com/ambient-memory/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Source: &lt;a href="https://github.com/JarvisOsorio/ambient-memory" rel="noopener noreferrer"&gt;https://github.com/JarvisOsorio/ambient-memory&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;License: MIT&lt;/li&gt;
&lt;li&gt;Version in the repo: 0.3.0&lt;/li&gt;
&lt;li&gt;Language: TypeScript, Node.js 20+ (22+ preferred)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I'm Jarvis Osorio (Jarvis Nelson Osorio), a technology entrepreneur and AI innovator based in Miami Beach, Florida.&lt;/p&gt;

&lt;h2&gt;
  
  
  The problem I wanted to solve
&lt;/h2&gt;

&lt;p&gt;Most "memory" for coding agents is either a Markdown vault you have to maintain, a cloud sync you have to trust, or a native SQLite addon that breaks the install on a clean machine. I wanted something that starts with one command, stays on disk, and speaks MCP.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Ambient Memory actually does
&lt;/h2&gt;

&lt;p&gt;A background daemon:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;watches Documents, Desktop, Downloads and the current working directory&lt;/li&gt;
&lt;li&gt;indexes screen sessions and browser history, with noise filters (Slack, Spotify and similar apps are ignored; focus churn is deduped)&lt;/li&gt;
&lt;li&gt;is the only process that opens the SQLite vault&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The CLI and the MCP server are thin clients over a local Unix socket (or a Windows named pipe). That design removes seal/unseal races from concurrent database opens.&lt;/p&gt;

&lt;p&gt;Storage uses Node's built-in &lt;code&gt;node:sqlite&lt;/code&gt; on Node 22+, with a &lt;code&gt;sql.js&lt;/code&gt; WASM fallback on Node 20. There is no &lt;code&gt;better-sqlite3&lt;/code&gt; and no &lt;code&gt;node-gyp&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Search is FTS5/BM25. Entity fields are encrypted with AES-256-GCM. When the daemon stops, the database is sealed to &lt;code&gt;memory.db.enc&lt;/code&gt; and the plaintext plus WAL/SHM files are removed. While the vault is unlocked, the FTS index is &lt;strong&gt;redacted plaintext&lt;/strong&gt; — an intentional search tradeoff, documented in &lt;a href="https://github.com/JarvisOsorio/ambient-memory/blob/main/SECURITY.md" rel="noopener noreferrer"&gt;SECURITY.md&lt;/a&gt;, not claimed as ciphertext. v0.3 has no telemetry and no network sync.&lt;/p&gt;

&lt;h2&gt;
  
  
  MCP tools
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;memory_search&lt;/code&gt; · &lt;code&gt;memory_recent&lt;/code&gt; · &lt;code&gt;memory_get&lt;/code&gt; · &lt;code&gt;memory_context&lt;/code&gt; · &lt;code&gt;memory_relate&lt;/code&gt; · &lt;code&gt;memory_status&lt;/code&gt;&lt;/p&gt;

&lt;p&gt;&lt;code&gt;memory_context&lt;/code&gt; returns a short "what was I doing" summary for the last N minutes, including a noise-control report.&lt;/p&gt;

&lt;h2&gt;
  
  
  Quick start
&lt;/h2&gt;

&lt;p&gt;As of 12 November 2026, &lt;code&gt;ambient-memory&lt;/code&gt; is not yet on the npm registry. Install from GitHub:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/JarvisOsorio/ambient-memory.git
&lt;span class="nb"&gt;cd&lt;/span&gt; ambient-memory
npm install
npm &lt;span class="nb"&gt;test&lt;/span&gt;
node ./bin/ambient.js init
node ./bin/ambient.js start --screen auto    &lt;span class="c"&gt;# headless: --screen simulated&lt;/span&gt;
node ./bin/ambient.js mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Point Cursor or Claude Desktop at the local binary:&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;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"ambient-memory"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"node"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"/ABS/PATH/ambient-memory/bin/ambient.js"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"mcp"&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;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The README also documents &lt;code&gt;npx -y ambient-memory&lt;/code&gt; as the intended path once the package is published to npm.&lt;/p&gt;

&lt;p&gt;Issues and contributions are welcome in the repo. If you try it, I'd like to know whether the daemon-owned vault and the honest FTS tradeoff match how you actually work with agents.&lt;/p&gt;

&lt;p&gt;— Jarvis Osorio&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>mcp</category>
      <category>memory</category>
    </item>
    <item>
      <title>Jarvis Nelson Osorio on Open-Weight Models: What a Full Release Changes for Builders</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:18:02 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-nelson-osorio-on-open-weight-models-what-a-full-release-changes-for-builders-3l0o</link>
      <guid>https://dev.to/jarvisosorio/jarvis-nelson-osorio-on-open-weight-models-what-a-full-release-changes-for-builders-3l0o</guid>
      <description>&lt;p&gt;I am &lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;, also known as &lt;strong&gt;Jarvis Osorio&lt;/strong&gt;, a technology entrepreneur and innovator based in Miami Beach, FL. When a serious lab releases the full weights of a flagship model under a permissive license, the story is not only "a new benchmark." The story is that more of the stack can be inspected, hosted, and adapted by people who do not sit inside that lab.&lt;/p&gt;

&lt;p&gt;I originally shared a technical walk-through of one such moment on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/deep-diving-zais-glm-52-jarvis-nelson-osorio-ramos-5us7f" rel="noopener noreferrer"&gt;Deep Diving into Z.ai’s GLM-5.2&lt;/a&gt;&lt;/em&gt;. I will not repeat a leaderboard here. Leaderboards move. The builder questions last.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why full weights are a different kind of news
&lt;/h2&gt;

&lt;p&gt;An API is a rental. You can do remarkable work with a rental, and I use hosted models constantly. You cannot, however, guarantee the same behavior next year, audit every internal change, or run the system in an environment you fully control. A full-weight release—especially under a license that allows real use—lets teams study the model, serve it themselves, and decide when to update.&lt;/p&gt;

&lt;p&gt;That matters for startups and for public institutions that cannot send every document to a distant endpoint. It also matters for researchers who want to understand failures instead of only reporting them.&lt;/p&gt;

&lt;h2&gt;
  
  
  Efficiency is part of the openness story
&lt;/h2&gt;

&lt;p&gt;The release I wrote about used a Mixture-of-Experts design: a very large total parameter count, with only a portion active on each pass, plus a long context window trained with coding-agent work in mind. The details belong to that model. The pattern is wider. Open does not have to mean "too heavy to run." Designers are working hard to make frontier-like reasoning cheaper to serve, including on hardware a serious team might actually own.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I tell teams in practice
&lt;/h2&gt;

&lt;p&gt;Do not switch your whole product because a release day was exciting. Run the same evaluations you already trust: your tasks, your languages, your latency and cost budget, your safety cases. Compare hosted and self-hosted paths honestly, including the people-cost of operations. If you do self-host, treat the serving layer as production infrastructure—access control, logging, update policy—not as a weekend experiment that accidentally became the company.&lt;/p&gt;

&lt;p&gt;Open weights are also an institutional opportunity. You can place the model next to your own constitution: tool allowlists, independent checks, records of what was asked and what was done. I develop that habit in &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;. Openness without governance is just a more portable black box.&lt;/p&gt;

&lt;p&gt;The good news is cultural as well as technical. A healthy ecosystem is not a single throne. It is many serious options, some rented and some owned, and builders who can move between them without panic.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;p&gt;This expands on ideas I originally shared on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/deep-diving-zais-glm-52-jarvis-nelson-osorio-ramos-5us7f" rel="noopener noreferrer"&gt;Deep Diving into Z.ai’s GLM-5.2&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More essays, books and profiles: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt; and the hub &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. For the institutional half of this argument, see &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More from Jarvis Osorio: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;https://jarvisosorio.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>ai</category>
      <category>llm</category>
      <category>startup</category>
    </item>
    <item>
      <title>Jarvis Osorio: Designing the Group Chat Between AI Agents</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:17:41 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-osorio-designing-the-group-chat-between-ai-agents-3p5c</link>
      <guid>https://dev.to/jarvisosorio/jarvis-osorio-designing-the-group-chat-between-ai-agents-3p5c</guid>
      <description>&lt;p&gt;I am &lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;, also known as &lt;strong&gt;Jarvis Osorio&lt;/strong&gt;—a technology entrepreneur and innovator based in Miami Beach, FL, working on AI systems, DAO institutions, books and art. I have joked, and I also mean it, that the only group chat I want to be in is the one with my AI agents. The punchline is not that people do not matter. It is that a working group of agents should not be designed like a noisy social channel.&lt;/p&gt;

&lt;p&gt;I originally shared this on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/only-group-chat-i-want-my-ai-agents-jarvis-nelson-osorio--1lqbf" rel="noopener noreferrer"&gt;The Only Group Chat I Want to Be in Is with My AI Agents&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why the metaphor matters
&lt;/h2&gt;

&lt;p&gt;Human teams spend a real share of their energy on coordination: context switching, status meetings, the slow recovery of attention after an interruption. Agents do not get tired, do not negotiate credit, and can pick up a repository or a log without a twenty-minute ramp. That is leverage, especially for a small team. It is also how people get tempted to drop structure and "just let them talk."&lt;/p&gt;

&lt;p&gt;A high-performing agent group is not a group chat in the casual sense. It is closer to a deterministic, event-driven network with roles.&lt;/p&gt;

&lt;h2&gt;
  
  
  Roles, not a crowd
&lt;/h2&gt;

&lt;p&gt;The pattern I come back to is simple:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;An &lt;strong&gt;orchestrator&lt;/strong&gt; takes a high-level goal, breaks it into work, and watches the path.&lt;/li&gt;
&lt;li&gt;An &lt;strong&gt;engineering&lt;/strong&gt; agent works inside a sandbox: code, tests, terminal, drafts.&lt;/li&gt;
&lt;li&gt;A &lt;strong&gt;QA and security&lt;/strong&gt; agent reviews what was produced before it gets near production.&lt;/li&gt;
&lt;li&gt;A &lt;strong&gt;research&lt;/strong&gt; agent gathers sources, live information and documentation, without pretending that a summary is a citation.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Protocols such as the Model Context Protocol let those roles share tools and files without a human acting as a copy-paste bridge. The human still belongs in the loop—just not as the clerk of every message. The human is the architect and the verifier, especially for steps that are hard to undo.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I refuse to automate away
&lt;/h2&gt;

&lt;p&gt;Vision, taste, and approval of irreversible actions stay with a person. Overnight refactors and draft pull requests can wait for morning review. Codebase sovereignty matters too: agents should run in environments you control, not in a fog that quietly trains on your intellectual property.&lt;/p&gt;

&lt;p&gt;This is the same separation I argue for in &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;. A swarm without roles is a committee with a faster mouth. A swarm with an orchestrator, a reviewer and a record is an institution.&lt;/p&gt;

&lt;p&gt;If you want a group chat with agents, design it as a workplace with a constitution: named roles, a sandbox, a reviewer who is not the author, and a human who can still say no.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;p&gt;This expands on ideas I originally shared on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/only-group-chat-i-want-my-ai-agents-jarvis-nelson-osorio--1lqbf" rel="noopener noreferrer"&gt;The Only Group Chat I Want to Be in Is with My AI Agents&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More essays, books and profiles: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt; and the hub &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. On institutions before unchecked power, see &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More from Jarvis Osorio: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;https://jarvisosorio.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>llm</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Jarvis Osorio on the Plain-Text Revolution: Why Markdown Still Powers AI Work</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:17:19 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-osorio-on-the-plain-text-revolution-why-markdown-still-powers-ai-work-4838</link>
      <guid>https://dev.to/jarvisosorio/jarvis-osorio-on-the-plain-text-revolution-why-markdown-still-powers-ai-work-4838</guid>
      <description>&lt;p&gt;I am &lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;, also known as &lt;strong&gt;Jarvis Osorio&lt;/strong&gt;, a technology entrepreneur and innovator based in Miami Beach, FL. Before I talk about agents, I want to talk about a quieter standard: plain text with just enough structure to be kind to humans and to machines. That standard is Markdown, and it is having a second life in AI systems.&lt;/p&gt;

&lt;p&gt;I originally shared this history and argument on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/plain-text-revolution-brief-history-markdown-why-powers-osorio--rposf" rel="noopener noreferrer"&gt;The Plain-Text Revolution: A Brief History of Markdown and Why It Powers Modern Software and AI&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  A 2004 writing hack that refused to die
&lt;/h2&gt;

&lt;p&gt;In the early web you had an ugly choice. Raw HTML was heavy with tags. Rich editors locked writing into brittle, often binary formats that fought version control and rendered differently everywhere. Plain &lt;code&gt;.txt&lt;/code&gt; files traveled well but had no hierarchy—no honest heading, list, or link.&lt;/p&gt;

&lt;p&gt;In 2004, John Gruber, working with Aaron Swartz, proposed Markdown: a lightweight markup you can read as ordinary text, with simple markers for headings, emphasis and lists, that can be turned cleanly into HTML. It was a writing tool. Then it became the language of software documentation, because Git can show a useful diff of a &lt;code&gt;.md&lt;/code&gt; file the way it cannot show a useful diff of a sealed document.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why language models like the same files
&lt;/h2&gt;

&lt;p&gt;Agents and retrieval pipelines thrive on text they can parse without a ceremony. Markdown is compact, so you spend fewer tokens on decoration. It has a visible skeleton, so a model can find the section that matters. It can live next to code in the same repository, which means your product facts and your product source can share a history.&lt;/p&gt;

&lt;p&gt;When teams keep the source of truth in slide decks and locked files, every agent has to guess. When the source of truth is a versioned Markdown corpus—product requirements, runbooks, decision records—the agent has a map. That is not a small convenience. It is how you ground a system in something other than a confident improvisation.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I treat it in an agentic workflow
&lt;/h2&gt;

&lt;p&gt;I want a living map between research, design and code, and I often want that map to be a plain file the whole team can read. The same instinct shows up in the "design.md" bridge I have described in other writing: give the agent a document it can revisit, not only a chat it will forget. Persistence favors formats that survive a handoff.&lt;/p&gt;

&lt;p&gt;This is also a cultural choice. Fancy tools come and go. A folder of Markdown, in a repository your team actually owns, is a kind of institutional memory. It is closer to a constitution than to a mood. That is why it sits comfortably beside the arguments in &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;: if the rules of a system cannot be read, they cannot be contested.&lt;/p&gt;

&lt;p&gt;If you are building with language models this year, do not start with a new dashboard. Start by putting the sentences you already trust into files a human can read, a git history can track, and an agent can retrieve without getting lost in formatting.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;p&gt;This expands on ideas I originally shared on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/plain-text-revolution-brief-history-markdown-why-powers-osorio--rposf" rel="noopener noreferrer"&gt;The Plain-Text Revolution: A Brief History of Markdown and Why It Powers Modern Software and AI&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More essays, books and profiles: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt; and the hub &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. For how institutions need readable rules, see &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More from Jarvis Osorio: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;https://jarvisosorio.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>markdown</category>
      <category>ai</category>
      <category>llm</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Jarvis Nelson Osorio: Why Finance AI Agents Need an Independent Judge</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:16:55 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-nelson-osorio-why-finance-ai-agents-need-an-independent-judge-3be</link>
      <guid>https://dev.to/jarvisosorio/jarvis-nelson-osorio-why-finance-ai-agents-need-an-independent-judge-3be</guid>
      <description>&lt;p&gt;I am &lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;, also known as &lt;strong&gt;Jarvis Osorio&lt;/strong&gt;, a technology entrepreneur and innovator based in Miami Beach, FL. I work on AI systems, DAO design, writing and art. As soon as an agent can touch money—rebalance a book, prepare a transfer, reconcile an invoice, or construct a transaction—the most dangerous phrase in the architecture is "it checked itself."&lt;/p&gt;

&lt;p&gt;I originally shared this argument on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/why-finance-ai-agents-need-independent-validation-inside-osorio--8ynjf" rel="noopener noreferrer"&gt;Why Finance AI Agents Need Independent Validation: Inside Screwtape.ai&lt;/a&gt;&lt;/em&gt;. The product landscape will move. The design rule should not: the process that proposes a financial action must not be the only process that blesses it.&lt;/p&gt;

&lt;h2&gt;
  
  
  The self-grading problem, with a price tag
&lt;/h2&gt;

&lt;p&gt;In many multi-agent demos, one model writes the plan and the same family of model "reviews" it. That feels like governance. In numerical and contractual work it is often an echo. If the first pass misreads a margin, a date, or a clause, the second pass is statistically likely to find the error reasonable. In a chatbot, that is an awkward paragraph. In a ledger, it is capital.&lt;/p&gt;

&lt;p&gt;Finance does not forgive a confident hallucination the way a brainstorming session can. A single unverified write can be very expensive. So the standard I argue for is simple and strict: before an agent executes a trade, issues an invoice, or updates a balance, an independent layer should return a clear verdict—pass, fail, or retry—against constraints you actually wrote down.&lt;/p&gt;

&lt;h2&gt;
  
  
  Consensus can be a false friend
&lt;/h2&gt;

&lt;p&gt;Teams sometimes add more agents and take comfort when they agree. I have also written about the opposite case: several agents can share a flawed assumption and call it confidence. In a financial swarm, agreement should be scored, and "we all said the same number" should raise a question: did they share a bad input? An independent check that flags false-consensus risk is more useful than another teammate trained the same way.&lt;/p&gt;

&lt;h2&gt;
  
  
  Receipts, not vibes
&lt;/h2&gt;

&lt;p&gt;If a regulator, a partner, or your future self asks why an autonomous system moved money, "the model thought it was fine" is not a record. A signed, inspectable receipt of the checks performed &lt;em&gt;before&lt;/em&gt; execution is. This is institutional design, the same family of ideas I develop in &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;: visible rules, separated powers, an audit trail that does not depend on anyone's memory.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "independent" has to mean
&lt;/h2&gt;

&lt;p&gt;Independence is not a second tab with the same prompt. It is a different evaluator, a deterministic test, or an external service that cannot be talked out of a rule. It sits beside the executing agent, not inside its self-talk. It should be allowed to say no.&lt;/p&gt;

&lt;p&gt;You can build that layer yourself or use a dedicated judge service; the original LinkedIn piece looks at one such pattern. Either way, do not let a finance agent hold the pen and the gavel. Let it draft. Let something else, with a record, decide whether the draft may become a transaction.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;p&gt;This expands on ideas I originally shared on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/why-finance-ai-agents-need-independent-validation-inside-osorio--8ynjf" rel="noopener noreferrer"&gt;Why Finance AI Agents Need Independent Validation: Inside Screwtape.ai&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More essays, books and profiles: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt; and the hub &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. The same separation of powers is the subject of &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More from Jarvis Osorio: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;https://jarvisosorio.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>fintech</category>
      <category>llm</category>
    </item>
    <item>
      <title>Jarvis Osorio: Why AI Agents Loop, What It Costs, and How Architecture Stops It</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:16:32 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-osorio-why-ai-agents-loop-what-it-costs-and-how-architecture-stops-it-1blc</link>
      <guid>https://dev.to/jarvisosorio/jarvis-osorio-why-ai-agents-loop-what-it-costs-and-how-architecture-stops-it-1blc</guid>
      <description>&lt;p&gt;I am &lt;strong&gt;Jarvis Osorio&lt;/strong&gt; (&lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;), a technology entrepreneur and innovator based in Miami Beach, FL. I write about AI systems, DAO institutions, books and art. One failure I now see everywhere, once teams leave the chat box and let agents actually &lt;em&gt;do&lt;/em&gt; work, is the loop: the same tool call, the same error, the same retry, the same bill.&lt;/p&gt;

&lt;p&gt;I originally shared this diagnosis on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/why-your-ai-agent-looping-root-cause-economic-cost-fixes-osorio--grrwf" rel="noopener noreferrer"&gt;Why Is Your AI Agent Looping? The Root Cause, Economic Cost, and Architectural Fixes&lt;/a&gt;&lt;/em&gt;. The short version is unsentimental. Looping is rarely a mysterious glitch. It is usually an architecture symptom.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three ways a loop gets born
&lt;/h2&gt;

&lt;p&gt;The first is the &lt;strong&gt;self-grading trap&lt;/strong&gt;. The same model that takes the step is asked to judge whether the step worked. If it misunderstands its own mistake, it treats the failure as a formatting problem, tweaks a small parameter, and tries again. The echo chamber is polite, busy and wrong.&lt;/p&gt;

&lt;p&gt;The second is &lt;strong&gt;ambiguous stopping criteria&lt;/strong&gt;. Prompts such as "keep refining this until it is optimal" sound ambitious. They give the system no finish line. Language models work in probabilities, not hard gates. "Optimal" can move forever, and the meter keeps running.&lt;/p&gt;

&lt;p&gt;The third is &lt;strong&gt;tool-output misinterpretation&lt;/strong&gt;. A messy error, an empty list, or a confusing status code arrives. Instead of escalating or stopping, the agent "fixes" the input in a way that triggers the same failure downstream. The loop is a misunderstanding with a retry button.&lt;/p&gt;

&lt;p&gt;None of this requires a weak model. A capable model in an unconstrained loop can still burn tokens, hammer an API, and fail to deliver.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the loop actually costs
&lt;/h2&gt;

&lt;p&gt;The obvious cost is money: repeated calls, expanded context, idle compute. The quieter cost is trust. A team that has watched an agent spin overnight becomes cautious about giving it any real authority—the opposite of what you wanted when you automated the task. There is also a safety cost. A loop that retries a write, a payment or a message is not merely inefficient. It is a process without a circuit breaker.&lt;/p&gt;

&lt;h2&gt;
  
  
  Architecture, not pep talks
&lt;/h2&gt;

&lt;p&gt;The fix is not a cleverer pep prompt. It is a set of hard rails:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Cap iterations at the orchestrator, not inside the model's self-talk. A sub-task that may call a tool only a handful of times is already safer than an open &lt;code&gt;while&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Take evaluation away from the actor. A separate judge, or a deterministic test, should return a clear pass or fail before the next step is allowed.&lt;/li&gt;
&lt;li&gt;Prefer a state machine to a free-form conversation. Failure should lead to a named recovery or a stop, not another improvisation.&lt;/li&gt;
&lt;li&gt;Write playbooks for common errors: if this status arrives, wait, retry twice, then abort and notify a person.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I make the same argument in a different domain in &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;: institutions beat vibes. A DAO that lets the spender approve the spend is one failure mode. An agent that grades its own work is another.&lt;/p&gt;

&lt;p&gt;As models get faster, the bottleneck is no longer raw fluency. It is whether the system can stop, tell the truth about failure, and hand control back. Reliability will not come from asking the agent to try harder. It will come from constraints you are willing to write down.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;p&gt;This expands on ideas I originally shared on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/why-your-ai-agent-looping-root-cause-economic-cost-fixes-osorio--grrwf" rel="noopener noreferrer"&gt;Why Is Your AI Agent Looping? The Root Cause, Economic Cost, and Architectural Fixes&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More essays, books and profiles: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt; and the hub &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. For a parallel argument about systems that never refuse, see &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLVFB72B" rel="noopener noreferrer"&gt;The Compliant Mind&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More from Jarvis Osorio: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;https://jarvisosorio.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>llm</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Jarvis Osorio on Always-On AI Agents: What Changes When the Chat Window Never Closes</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:16:05 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-osorio-on-always-on-ai-agents-what-changes-when-the-chat-window-never-closes-2929</link>
      <guid>https://dev.to/jarvisosorio/jarvis-osorio-on-always-on-ai-agents-what-changes-when-the-chat-window-never-closes-2929</guid>
      <description>&lt;p&gt;I am &lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;, also known as &lt;strong&gt;Jarvis Osorio&lt;/strong&gt;—a technology entrepreneur and innovator, AI systems architect, DAO institutional designer, author and artist based in Miami Beach, FL. For years, working with a language model meant opening a box, typing a request, and watching the answer arrive. When you closed the laptop, the work stopped. That habit trained us to think of AI as a conversation. Always-on agents train us to think of AI as a colleague that keeps a desk even when we leave the room.&lt;/p&gt;

&lt;p&gt;This is the shift I originally shared on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/openai-dots-arrival-always-on-persistent-ai-agents-osorio-ramos-j1uuf" rel="noopener noreferrer"&gt;OPENAI DOTS: THE ARRIVAL OF ALWAYS-ON, PERSISTENT AI AGENTS&lt;/a&gt;&lt;/em&gt;. Vendor names will change. The architectural change will not: a persistent runtime, a memory that compounds, and a machine that can keep acting after you log off.&lt;/p&gt;

&lt;h2&gt;
  
  
  From a reply to a runtime
&lt;/h2&gt;

&lt;p&gt;A chat session is ephemeral. It holds a thread, then it fades. An always-on agent is closer to a small computer that happens to reason. In the pattern I described, the agent is given its own isolated cloud machine and a sandboxed browser, so it can use tools and interfaces the way a person would—not only through a neat API, but also by looking at a screen when no API exists. You can assign a long audit, a refactor, or a watchful monitoring job, disconnect, and come back to logs, scratch files and a trail of what already happened.&lt;/p&gt;

&lt;p&gt;That sounds like a gift. It is also a constitution problem. A system that works while you sleep can also spend, message, edit or delete while you sleep. Persistence without rules is just unsupervised motion.&lt;/p&gt;

&lt;h2&gt;
  
  
  Memory that compounds
&lt;/h2&gt;

&lt;p&gt;Ordinary chat memory is a pile of facts. Longitudinal memory is a working model of how &lt;em&gt;you&lt;/em&gt; like work to be done: tone, quality bar, the projects already in flight. Feedback makes the next run sharper. That is useful. It is also why you should be deliberate about what the agent is allowed to remember and which tools it may touch on day one. Connect only the calendar, repository or inbox the first job actually needs. Extra connectors are extra blast radius.&lt;/p&gt;

&lt;h2&gt;
  
  
  Watch the work, do not only read the summary
&lt;/h2&gt;

&lt;p&gt;Transparency is not a luxury here. An activity view—work in progress, scheduled, and completed—lets you open the agent's environment and watch it navigate or run code, then pause it. I treat that the way I treat a treasury dashboard in a DAO: if you cannot see the action, you do not really govern it. The same instinct sits at the center of &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;: write the rules before you hand over power, separate the actor from the approver, and keep a record someone else can inspect.&lt;/p&gt;

&lt;h2&gt;
  
  
  What actually changes in the job
&lt;/h2&gt;

&lt;p&gt;When the window never closes, your job moves again. Less time pasting prompts. More time writing standing orders: which actions may run alone, which need a human yes, and what "done" looks like so the agent can stop. Always-on execution is a feature. Unbounded authority is not.&lt;/p&gt;

&lt;p&gt;If you are excited to give an agent a desk of its own, start with the smallest desk, the fewest keys, and a door you can still lock.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where to go next
&lt;/h2&gt;

&lt;p&gt;This expands on ideas I originally shared on LinkedIn in &lt;em&gt;&lt;a href="https://www.linkedin.com/pulse/openai-dots-arrival-always-on-persistent-ai-agents-osorio-ramos-j1uuf" rel="noopener noreferrer"&gt;OPENAI DOTS: THE ARRIVAL OF ALWAYS-ON, PERSISTENT AI AGENTS&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More essays, books and profiles: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt; and the hub &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. For the human side of systems that never stop agreeing, see &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLVFB72B" rel="noopener noreferrer"&gt;The Compliant Mind&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;More from Jarvis Osorio: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;https://jarvisosorio.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>llm</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Jarvis Nelson Osorio on System-One Models: Why Micro-Decisions Under 100ms Change Product Design</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:09:22 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-nelson-osorio-on-system-one-models-why-micro-decisions-under-100ms-change-product-design-39h9</link>
      <guid>https://dev.to/jarvisosorio/jarvis-nelson-osorio-on-system-one-models-why-micro-decisions-under-100ms-change-product-design-39h9</guid>
      <description>&lt;p&gt;&lt;em&gt;Canonical: &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-on-System-One-Models-Why-Micro-Decisions-Under-100ms-Change-Product-Design-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio on System-One Models: Why Micro-Decisions Under 100ms Change Product Design&lt;/a&gt; — also at &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt;. Originally published by Jarvis Nelson Osorio (Jarvis Osorio).&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Most public conversation about AI still sounds like a conversation about essays. Models write pages, summarize meetings, generate slides. I am interested in a quieter layer: the &lt;strong&gt;micro-decision&lt;/strong&gt;, the tiny judgment a product must make before the user even notices a pause. I am &lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;, also &lt;strong&gt;Jarvis Osorio&lt;/strong&gt;, a technology entrepreneur and innovator based in Miami Beach, FL, working in AI systems, DAO design, writing and art.&lt;/p&gt;

&lt;p&gt;I have been writing about &lt;strong&gt;system-one style models&lt;/strong&gt; — specialists aimed at very fast, narrow choices — because many product moments cannot wait for a long generated paragraph. Some of the industry discussion points to budgets like a tenth of a second. I treat that number as a design ambition, not as a laboratory claim I am here to certify. The point is the architecture: when speed becomes the feature, the model you choose and the way you wire it both change.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two speeds of mind, two speeds of product
&lt;/h2&gt;

&lt;p&gt;A useful metaphor, borrowed from everyday psychology rather than from a vendor brochure, is the difference between a fast, automatic judgment and a slow, deliberate one. Products need both.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Fast path.&lt;/strong&gt; Is this message spam? Is this code completion even plausible? Should we show a warning? Which tool should we call? May this draft be sent to a child-facing surface?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Slow path.&lt;/strong&gt; Write the analysis. Compare the legal language. Plan the migration. Sit with a hard tradeoff.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you force the fast path through a large, distant, conversational model, you get two problems at once: the user feels lag, and you pay for a monologue when you needed a nod. If you force the slow path through a tiny classifier, you get cheap confidence and thin reasoning.&lt;/p&gt;

&lt;p&gt;System-one models, as I use the phrase, are for the fast path. They should be small, specialized, close to the product, and honest about their limits. When they are unsure, they escalate.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why "under a blink" is a product decision
&lt;/h2&gt;

&lt;p&gt;Latency is not only comfort. It changes behavior. Editors, games, trading tools, classroom apps, and agent runtimes all have moments where a delay teaches the user to skip the check. If a safety or quality gate is slow, people will work around it. If it is fast enough to feel invisible, it can actually be used.&lt;/p&gt;

&lt;p&gt;That is why local and high-speed agents belong in the same conversation. A model that lives nearer the user can make a micro-decision without a transoceanic round trip — something teams in Latin America feel more often than keynotes admit. Privacy rides along: a fast local check can happen without shipping a full document to a public endpoint.&lt;/p&gt;

&lt;h2&gt;
  
  
  Design rules I actually use
&lt;/h2&gt;

&lt;p&gt;When I sketch a system with a fast layer, I write the rules down:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Name the micro-decisions.&lt;/strong&gt; If you cannot list them, you will accidentally use a giant model as a reflex.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Give each one a budget&lt;/strong&gt; — time, cost, and what "I don't know" looks like.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Escalate on uncertainty.&lt;/strong&gt; The fast model is allowed to pass. It is not allowed to bluff.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep the fast model boring.&lt;/strong&gt; Boredom is testable. A witty micro-decider is a side effect, not a goal.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Log the decision.&lt;/strong&gt; Even a 40-millisecond choice should be reconstructible after an incident.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Those rules are institutional. They rhyme with the way I think about treasuries and agent permissions in &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;: limit authority, record the act, do not confuse charm with control.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where yes-machines hide in the fast path
&lt;/h2&gt;

&lt;p&gt;A micro-decision can be a yes-machine too. "Always allow" is extremely fast. "Always agree with the user's implied wish" is also fast. Speed without a concept of refusal is just a quicker way to go wrong. &lt;em&gt;The Compliant Mind&lt;/em&gt; is mostly about slower, conversational systems, but the lesson travels: some moments should still be allowed to say no, even in under a blink.&lt;/p&gt;

&lt;p&gt;The hopeful version is this: once the fast path is reliable, the slow path can be more humane. People get immediate help on the small things and reserved, careful attention on the large ones. That is a better product than a single model pretending to be both a reflex and a philosopher.&lt;/p&gt;

&lt;p&gt;I write and publish from &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt;. A longer list of essays, books and videos is on the hub &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. For the deliberate side of judgment, read &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLVFB72B" rel="noopener noreferrer"&gt;The Compliant Mind&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;The next interesting products will not only generate more text. They will decide small things well, quickly, and in a way we can still explain.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>startup</category>
      <category>latency</category>
      <category>agents</category>
    </item>
    <item>
      <title>Jarvis Nelson Osorio: Building a Tech Company for Latin America</title>
      <dc:creator>Jarvis Nelson Osorio</dc:creator>
      <pubDate>Tue, 10 Nov 2026 02:07:09 +0000</pubDate>
      <link>https://dev.to/jarvisosorio/jarvis-nelson-osorio-building-a-tech-company-from-chile-and-latin-america-2k53</link>
      <guid>https://dev.to/jarvisosorio/jarvis-nelson-osorio-building-a-tech-company-from-chile-and-latin-america-2k53</guid>
      <description>&lt;p&gt;&lt;em&gt;Also at &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt;. Originally published by Jarvis Nelson Osorio (Jarvis Osorio).&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;I am based in Miami Beach, Florida, building for Latin American markets. That is not a branding footnote. It is the operating condition of my work. I am &lt;strong&gt;Jarvis Nelson Osorio&lt;/strong&gt;, also &lt;strong&gt;Jarvis Osorio&lt;/strong&gt; — a technology entrepreneur and innovator, AI and systems architect, DAO institutional designer, author and artist. When people still talk as if serious technology can only be assembled in a handful of coastal cities, I treat that as an old map. The interesting question is how to build well for the region.&lt;/p&gt;

&lt;h2&gt;
  
  
  What building for Latin America actually changes
&lt;/h2&gt;

&lt;p&gt;Distance is real. Some meetings happen when it is late. Some services assume a card, a courier, or a legal template that does not travel cleanly. Latency to a far-away model endpoint is not an abstract benchmark; it is something a team in Latin America feels in the editor. Currency and infrastructure costs shape which tools you keep on a monthly bill.&lt;/p&gt;

&lt;p&gt;None of that is a reason to postpone ambition. It is a reason to be deliberate:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Prefer architectures that can run closer to the team when privacy or speed matters — local and small models included.&lt;/li&gt;
&lt;li&gt;Write institutions early. If your collaborators are in more than one country, informal hallway trust will not scale. This is one reason I care about DAO design and about the ideas in &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;: rules, records, and clear authority travel better than vibes.&lt;/li&gt;
&lt;li&gt;Build for Spanish and Portuguese as first-class surfaces, not as afterthoughts. The region is not a translation layer on top of someone else's product.&lt;/li&gt;
&lt;li&gt;Use the time zone as an advantage when you can: overlap with both North America and Europe is a feature if you plan the day instead of apologizing for it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;South Florida, Spain, and other markets can be places you serve. They are not a requirement for a "real" headquarters. I am based in Miami Beach, Florida, building for Latin American markets.&lt;/p&gt;

&lt;h2&gt;
  
  
  Talent and taste are already here
&lt;/h2&gt;

&lt;p&gt;Latin America has operators who have worked around constraints for years. That experience shows up as resourcefulness, but it should also show up as taste: clear product judgment, respect for users' data, and an allergy to theater. I would rather hire someone who can explain a decision than someone who can recite a trend.&lt;/p&gt;

&lt;p&gt;There is a cultural asset that is easy to miss if you only import playbooks. Families, schools, and firms in this region often have a sharp sense of when a system is performing friendliness instead of being useful. That instinct is valuable in the age of yes-machines. &lt;em&gt;The Compliant Mind&lt;/em&gt; was written with those household and classroom moments in mind, not only with a laboratory audience.&lt;/p&gt;

&lt;h2&gt;
  
  
  Institutions beat imported slogans
&lt;/h2&gt;

&lt;p&gt;A payment, a grant, a token launch, or an AI agent with keys will test whether you actually have governance. Copying the vocabulary of a distant ecosystem — "community," "decentralization," "agentic" — without the machinery is how projects become confusing to the very people they hoped to include.&lt;/p&gt;

&lt;p&gt;What travels:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A written constitution for how money and permissions move.&lt;/li&gt;
&lt;li&gt;Public, boring reports.&lt;/li&gt;
&lt;li&gt;Separate people (or systems) for proposing and approving irreversible steps.&lt;/li&gt;
&lt;li&gt;Product defaults that can refuse, delay, or ask for a human.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;What does not travel well: charisma as a substitute for process, or the belief that being far from a traditional tech capital means you can skip the unglamorous parts.&lt;/p&gt;

&lt;h2&gt;
  
  
  A practical posture for founders here
&lt;/h2&gt;

&lt;p&gt;If I were sitting with a founder in Santiago, Lima, Buenos Aires or Montevideo this week, I would suggest a compact posture:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Pick a narrow product that you can actually support across time zones.&lt;/li&gt;
&lt;li&gt;Decide, in writing, where data lives and which models are allowed to see it.&lt;/li&gt;
&lt;li&gt;Make Spanish (and Portuguese, if relevant) part of the design, not a patch.&lt;/li&gt;
&lt;li&gt;Keep burn and tool sprawl visible; distance makes quiet subscriptions easier to ignore.&lt;/li&gt;
&lt;li&gt;Publish something you are willing to put your name on. Reputation compounds slowly and is worth more when you are not standing in the middle of someone else's scene.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I also keep making visual work — explosive visual representation — because a region this intense should not only speak in muted product screenshots. The art is part of the same practice: make pressure visible, then design with honesty.&lt;/p&gt;

&lt;p&gt;I am optimistic about building for this region. The tools have moved. The remaining work is institutional and cultural, and those are things we can author ourselves.&lt;/p&gt;

&lt;p&gt;Essays, books and profiles: &lt;a href="https://jarvisosorio.com" rel="noopener noreferrer"&gt;jarvisosorio.com&lt;/a&gt; · &lt;a href="https://telegra.ph/Jarvis-Nelson-Osorio-Jarvis-Osorio-articles-books-videos-and-profiles-11-10" rel="noopener noreferrer"&gt;Jarvis Nelson Osorio — articles, books, videos and profiles&lt;/a&gt;. A good companion for the institutional side is &lt;em&gt;&lt;a href="https://www.amazon.com/dp/B0HLWZVBCK" rel="noopener noreferrer"&gt;Breaking the Centralized Order&lt;/a&gt;&lt;/em&gt;.&lt;/p&gt;

</description>
      <category>startup</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
