AI Insights · Oct 23, 2026
Briefing, 23 October 2026: Agent Skills go to an open foundation
Anthropic gave the Agent Skills standard to the Agentic AI Foundation. Maintainers from Anthropic, Google and OpenAI propose versioned releases and security guidance.
What changed
- Agent Skills left vendor stewardship. Anthropic donated the standard to the Agentic AI Foundation on 2026-10-22, with maintainers from Anthropic, Google and OpenAI proposing versioned releases, a proposal process and security guidance. The "no evidence yet" and "remain vendor-stewarded" lines in two findings are now corrected.
- Claude Code mods are in-process and unsandboxed. From v2.1.287 they are on by default, run with the user's permissions, can approve tool calls and message other sessions, and sit outside the Bash sandbox.
- Unattended runs end in human review. Google's 11-hour Gemini 4 Argon run on ADK shipped 2 of 6 proposals after review; DoorDash's flag-cleanup agents got 45 of 50 PRs approved at $4.79 each, one git worktree per flag.
- Gemini 4 Argon's price is introductory. $2/$10 at launch, $4/$20 after, which is Opus 5.5 parity.
- Harness value depends on the model tier. In a 24-configuration vulnerability study a heavy prompt lifted smaller models by 20-86% and the frontier model by 2%. PAUSE gets SWE-bench gains from one plan at the first edit, but only with small executors.
- Plain similarity RAG misses most multi-fact evidence. Any-gold coverage 82%, all-gold 19% on a 50-query pilot.
Worth reading in full
- Claude Code mods overview, section "Decide whether to trust a mod": https://code.claude.com/docs/en/plugins/mods/overview
- Google, overnight Argon run on ADK, the prompt and the review outcome: https://developers.googleblog.com/optimizing-a-200000-line-python-codebase-overnight-with-gemini-4-argon/
- The Bottleneck Is Retrieval, Not Reading (2610.26951), for the evidence-based-generation metric.
AI-assisted briefing: compiled with AI from the week's papers, posts and releases, with links to each source.