---
title: "Stale docs, fresh agent mistakes"
newsletter: "User Community"
date: 2026-11-19
source: https://aaif.live/newsletters/usercommunity/2026-11-19-stale-docs-fresh-agent-mistakes
---

# Stale docs, fresh agent mistakes

*Plus… PRs prompted by production errors, 8,000 tools on demand, and AI malware on Android TV boxes.*

*User Community — Agentic AI Foundation, 2026-11-19*

Hard to describe the relief of realizing this warning wasn't about coffee [https://calif.io/java4shell/].

## Gems

## Community discussion & show & tell

## AAIF LUNCH AND LEARN

Friday, November 13 | 9 AM PT / 6 PM CET

We're opening up Session 31 for a community-led discussion. Bring along a coding agent tool you've been trying, a workflow you'd like to share, a challenge you've run into, or even a quick demo. There's no fixed agenda or formal presentation, just a chance to swap ideas, ask questions, and help shape future sessions.

Join the session here [https://home.mlops.community/home/events/coding-agents-lunch-and-learn-session-31-community-discussion-and-show-and-tell-fuqon8t7gj].

Check out past sessions here [https://www.youtube.com/watch?v=8U7eLahK9hE].

## What’s happening across the chapters

## LOCAL ORGANIZERS

FROM THE AAIF MUNICH COMMUNITY

The AAIF Munich community recently held its 29th meetup, with Wilson Cely [https://www.linkedin.com/in/wilsoncely/] introducing the foundation and ways to get involved.

In The Infrastructure Factory [https://home.mlops.community/public/videos/the-infrastructure-factory-agents-write-the-code-who-builds-the-place-it-runs], Max Körbächer [https://www.linkedin.com/in/maxkoerbaecher/] explored how teams can give AI agents secure, policy-compliant environments to run their code without lengthy approval processes.

Svetozar Miučin [https://www.linkedin.com/in/svetozar-miucin/] took a different approach in Two and a half wolves [https://home.mlops.community/public/videos/two-and-a-half-wolves], looking at how software engineers can balance building quickly, maintaining technical expertise, and deciding what's worth building in the first place.

Thanks to local organizers Wilson Cely, Marina Bevacqua [https://www.linkedin.com/in/marina-bevacqua/], and Jitendra Kasaudhan [https://www.linkedin.com/in/jitendra-kasaudhan-3298b632/] for bringing the community together.

You can find your upcoming local meetups here [https://aaif.io/events?tab=community].

## Why Stripe cut 500,000 internal docs down to 5,000

At one company, the top search result for new-engineer onboarding was an old intern doc marked “do not use.” Most internal doc collections are 95% clutter, and agents now read every page.

 * Agents rely on training data for speed and skip current pages, so they kept recommending competitors after Vercel added Docker support.

 * Agents want complete markdown, while people want diagrams, workflows, and checklists.

 * Timestamped decision records and named doc owners stop teams from rearguing old choices. One team spent weeks on an incident after an agent trusted a stale doc.

Teams that cut their docs down to a small set with named owners, which agents can keep in sync with the code, make documentation part of reliability work.

[https://podcasts.apple.com/gb/podcast/why-stripe-cut-500-000-internal-docs-down-to-5-000/id1505372978?i=1000801637091](https://podcasts.apple.com/gb/podcast/why-stripe-cut-500-000-internal-docs-down-to-5-000/id1505372978?i=1000801637091)

[https://home.mlops.community/home/videos/why-stripe-cut-500000-internal-docs-down-to-5000](https://home.mlops.community/home/videos/why-stripe-cut-500000-internal-docs-down-to-5000)

[https://open.spotify.com/episode/4mYtkYTUpDplhJGioR8xv6?si=okBQrWrnT1yW2hdqQTZENA](https://open.spotify.com/episode/4mYtkYTUpDplhJGioR8xv6?si=okBQrWrnT1yW2hdqQTZENA)

## What happens when engineers stop reading their own code?

Ten occurrences of the same error in an hour can prompt Miro’s internal agent to propose a fix. But getting useful PRs merged takes more than fast code generation.

 * Its AI gateway routes across cloud providers, models, and regions while respecting token quotas, costs, and European-only model requirements.

 * Cloudflare’s agent-facing engineering codex records approved tools, exceptions, and company practices to give coding agents more relevant context.

 * Miro combines log-driven PRs, automated reviews, and engineer feedback to distinguish relevant fixes from unwanted changes.

Reliable agent workflows depend on routing rules, shared engineering context, and feedback grounded in what engineers accept.

[https://podcasts.apple.com/gb/podcast/multi-cloud-llm-routing-and-context-aware-coding-agents/id1505372978?i=1000802313650](https://podcasts.apple.com/gb/podcast/multi-cloud-llm-routing-and-context-aware-coding-agents/id1505372978?i=1000802313650)

[https://home.mlops.community/home/videos/what-happens-when-engineers-stop-reading-their-own-code](https://home.mlops.community/home/videos/what-happens-when-engineers-stop-reading-their-own-code)

[https://open.spotify.com/episode/4NMBt9U5JLkolFcgEQLnIf?si=MjACBCIEQVmwoSOfol2cnw](https://open.spotify.com/episode/4NMBt9U5JLkolFcgEQLnIf?si=MjACBCIEQVmwoSOfol2cnw)

## Could your Android TV box run AI malware?

In an Anthropic lab test, small models running on Android TV boxes with open debug ports copied themselves to another device on the same network. The models were slow, and no attack like this has been reported outside a lab yet.

 * Malware that rewrites its own code changes its fingerprint each time, so signature-based antivirus can't match it.

 * In 2025, Google cut off Russian malware that called models hosted on Google Cloud. Malware that uses a local model already installed on the machine is harder to stop.

 * MCP tools can change their definitions after install. Setting an allow-list of tools when the connection is made prevents that.

Once malware can rewrite itself, defenders get more protection from network egress controls and tool allow-lists than from fingerprint matching.

[Video](https://home.mlops.community/home/videos/could-your-android-tv-box-run-ai-malware)

## How to give an agent 8,000 tools without loading 8,000 tools

An agent can fail a multi-step task because one tool is missing, even when thousands of others are available. Dynamic retrieval expands tool access without filling the context window or changing cached prompts.

 * A fixed interface retrieves relevant schemas on demand, using a two-stage retriever and reranker that responds in roughly 0.3 seconds.

 * On ToolRet, every required tool appeared in the top ten for 45.2% of queries, versus 28.8% with BM25.

 * Input tokens fell 17% with Codex and 40% with Claude Code, although fixed lists remained cheaper below roughly 25–28 tools.

The practical test is complete tool coverage and session cost on real workloads.

[Read the blog](https://aaif.io/blog/how-to-give-an-agent-8-000-tools-without-loading-8-000-tools)

## Why are people excited about Jev?

Agents authorize against MCP servers with OAuth, which was designed to stop people handing passwords to third-party apps. In this catch-up, they worry less about rogue agents than about the access controls around them.

 * With skills over MCP, teams could subscribe to vendor skills that stay current, though agents can do worse with several skills loaded.

 * A support agent could call Jev, a fast classifier, on each turn and bring in a human before customers type HUMAN.

 * Web MCP builders at AGNTCon + MCPCon in Amsterdam suggested exposing just a site’s top four user journeys as tools.

Teams adding agent surfaces like Web MCP take on new access control work with each one.

[Video](https://home.mlops.community/home/videos/why-are-people-excited-about-jev)

## How to move a skill from one coding agent to another

Same pull request, same migration-note skill, different results. The repo_search tool works in one coding agent but doesn’t exist in another, and although Agent Skills provides a shared format, hosts still differ in what they can discover, access, and execute.

 * Replace tool-specific commands with evidence requirements, such as inspecting the diff, finding API callers, and verifying search matches.

 * Validate SKILL.md metadata and directory names with skills-ref validate, then check that bundled files and dependencies are present in the new agent.

 * Test the same PR in the new agent for explicit and automatic loading, access to known callers, and accurate before-and-after examples.

A skill moves cleanly when each agent can back its migration note with evidence from the code, and a known caller missing from the note shows where to look next.

[Read the blog](https://aaif.io/blog/how-to-move-a-skill-from-one-coding-agent-to-another)

## Total cost of AI over value realized

In one enterprise audit, direct model consumption accounted for roughly a quarter of total AI spending. That leaves teams measuring more than token bills when assessing whether agents pay off.

 * Teams should count infrastructure, routing, observability, retries, idle compute, and human approvals in the cost of completing each task.

 * A cheaper model can cost more if its errors require supervision or hurt retention, so only acceptable outputs count toward value.

 * To show realized value, teams should distinguish capacity gains, augmentation, and automation, then connect claims to revenue, removed or avoided spending, or freed capital.

Comparing Total Cost of AI with verified Value Realized makes the trade-off between spending and outcomes visible.

[Read the blog](https://aaif.io/blog/total-cost-of-ai-over-value-realized)

## In-person events

## Come and connect

* Dhaka [https://luma.com/d819fdzd] - November 27

 * Pune [https://luma.com/gb5bwpen] - November 28

 * Helsinki [https://luma.com/eizq457i] - December 3

Find your city here [https://aaif.io/events?tab=community], or start a chapter if there isn't one yet.

## Virtual events

## Join from anywhere

* AAIF Reading Group: Scaling Discovery Requires Forgetting [https://home.mlops.community/home/events/aaif-reading-group-scaling-discovery-requires-forgetting-pg1zmi4qzx?agenda_day] - November 12

 * Lunch and Learn: Community discussion & show & tell [https://home.mlops.community/home/events/coding-agents-lunch-and-learn-session-31-community-discussion-and-show-and-tell-fuqon8t7gj] - November 13

---
Source: https://aaif.live/newsletters/usercommunity/2026-11-19-stale-docs-fresh-agent-mistakes
