AI Agent GitHub Digest — 2026-10-22

🌏 中文版

Today's Highlight

The wildest number today comes from a tool that spent half a year in obscurity — morluto/rea went from nobody-knows-it to 356,000 stars by letting agents reverse-engineer anything themselves. The same week, Microsoft and Cloudflare both open-sourced their own agent governance and audit tooling, echoing what we saw a few days ago with open-multi-agent: governance is moving from an afterthought to a first-class feature. And headroom attacks the agent cost problem from a different angle — compress what you're about to feed the LLM, and the savings show up directly on your token bill.

Trending Repos

morluto/rea ⭐ 356,198

GitHub　·　TypeScript　·　MIT

• What it is: An MCP server plus CLI that lets an agent figure out how a feature works without source code — hooking into Ghidra, IDA, and Hopper to analyze native binaries, plus static analysis for JavaScript/Electron apps, .NET assemblies, and websites. Everything runs locally.
• Why it matters: Created this past April, the repo now sits at 356.6k stars and 89.4k forks; observers watched it jump from a few thousand stars to hundreds of thousands within a single month in October (see the explainer video in the references). The pitch repackages reverse engineering — traditionally a niche security skill — as something an agent can just call: see a feature you like in someone else's app, point your agent at it, and get back the evidence of how it works plus a version you can build on.
• Tech stack: TypeScript core + MCP + wrappers around existing reverse-engineering tools like Ghidra, IDA, Hopper, and JADX
• Learning curve: Medium — npx rea-agents setup wires up your agent in one line, and static JavaScript analysis needs no extra engine, but native binary analysis requires an existing Ghidra, IDA, or Hopper install.

microsoft/agent-governance-toolkit ⭐ 6,452

GitHub　·　Python　·　MIT

• What it is: Microsoft's own AI agent governance toolkit — a policy enforcement engine, zero-trust identity, and execution sandboxing, with the project claiming coverage of all 10 OWASP Agentic Top 10 risks.
• Why it matters: A major vendor shipping agent governance as standard infrastructure rather than something you'd normally pay a consultancy for sits at the opposite end of the same story we saw with open-multi-agent (a community-built, telemetry-free governance runtime) — one is a lightweight DIY approach, the other a full vendor framework. Either way, agent governance has clearly stopped being a niche concern.
• Tech stack: Python + policy engine + zero-trust identity provider integration
• Learning curve: Medium — the pieces are all there, but you need to map your own agents' risk surface before you know which part to wire up.

cloudflare/security-audit-skill ⭐ 35,563

GitHub　·　JavaScript　·　MIT

• What it is: Cloudflare's own coding-agent skill that walks an agent through a multi-phase security audit, with every finding required to carry independently verifiable, machine-readable evidence.
• Why it matters: Unlike the usual "just ask the agent to find bugs" approach, breaking the audit into stages and requiring evidence for every finding cuts down on agents hallucinating vulnerabilities. Having Cloudflare maintain it officially is itself a signal that using agents as security auditors is moving from experimental to mainstream.
• Tech stack: Skill format for coding agents like Claude Code and Cursor (Markdown + a structured audit workflow)
• Learning curve: Low — install it like any other skill in an agent that supports the format.

headroomlabs-ai/headroom ⭐ 77,145

GitHub　·　Python　·　Apache-2.0

• What it is: Compresses tool output, logs, files, and RAG chunks before they reach the LLM — the project claims 20% fewer tokens for plain text and 60-95% for JSON, with no change in answer quality. Ships as a library, a proxy, or an MCP server.
• Why it matters: Most context-engineering discussion focuses on deciding what to send the model. Headroom goes after the other half of the problem — whether what you're already sending can be shrunk first. Offering three integration paths (library, proxy sitting between agent and tools, or MCP server) signals it wants to be a general-purpose layer rather than tied to one framework.
• Tech stack: Python + FastAPI (proxy mode) + MCP server, compatible with LangChain, Anthropic, and OpenAI
• Learning curve: Low — pip install the library or spin up the proxy in minutes; the MCP server path is just like installing any other MCP server.

Notable Releases

Pydantic AI v2.59.0

Release Notes

• Key changes: Fixed a high-severity security issue (GHSA-4whf-38hw-pxxh) where web_fetch_tool fetching attacker-controlled HTML with many elements could stall unrelated work in the same process during HTML conversion. Added DaytonaSandbox, letting Coder, Shell, FileSystem, and other harness capabilities run entirely inside a Daytona sandbox. Added ElevenLabsRealtimeModel for speech-to-speech realtime models.
• Breaking changes: Removed several already-retired Groq models (playai-tts and others). BackgroundTools results are now wrapped as tagged <tool_result> tool output instead of being passed as untagged user text.
• What this means for you: Upgrade to 2.59.0 (or 1.107.9) if your agent uses web_fetch_tool against untrusted pages, to close the DoS. If any code parses BackgroundTools output directly, update it to read the <tool_result> tag.

Mastra @mastra/core 1.79.0

Release Notes

• Key changes: Experimental MCP Events (SEP-3415 draft) let MCP servers publish event streams that clients can poll, keep open as a stream, or receive via signed webhooks. Observability gained aggregateSpans() for per-span analytics like p95 latency and token cost, plus new "time to first chunk" latency fields. Agent sessions are now addressed by sessionId instead of resourceId + scope, so a single resource can hold multiple concurrent sessions.
• Breaking changes: No formally flagged breaking change, but the session-addressing shift is worth checking if your code assumed a one-to-one mapping between a resourceId and a session.
• What this means for you: If you manage multiple conversations through Mastra's agent sessions, verify your code doesn't assume resourceId and session are one-to-one. MCP Events is experimental — try it on a non-critical path first.

Today's Takeaway

I used to assume open-source agent governance and audit tooling was mostly a patch job that smaller teams cobbled together to plug gaps. Seeing Microsoft and Cloudflare both ship this kind of tooling officially in the same week made it clear this has moved from "startup band-aid" to "big-vendor standard" — governance isn't something you bolt on after an incident anymore, it ships alongside the framework itself.

References

• morluto/rea on GitHub
• REA official site
• Reverse Engineer Anything (REA) EXPLAINED! — growth breakdown video
• microsoft/agent-governance-toolkit on GitHub
• cloudflare/security-audit-skill on GitHub
• headroomlabs-ai/headroom on GitHub
• Pydantic AI v2.59.0 release notes
• Pydantic AI GHSA-4whf-38hw-pxxh security advisory
• Mastra @mastra/core@1.79.0 release notes