On October 22, 2026, Anthropic handed the Agent Skills open standard — along with its name, logo, GitHub organisation and the agentskills.io domain — to the Agentic AI Foundation. It is the second major protocol Anthropic has donated to a Linux Foundation subsidiary in ten months. The first, in December 2025, was the Model Context Protocol. The company co-founded the AAIF last year alongside Block and OpenAI; it has now contributed two of its specifications to a foundation it helped stand up.

Anthropic says the standard has been integrated by more than 70 agent platforms — a list that includes ChatGPT, Cursor, GitHub Copilot and VS Code — with roughly one million skills published and nearly 280 million installs. The lead maintainers going forward are David Soria Parra and Mahesh Murag. Engineers from Google and OpenAI are also named as maintainers, "in their individual capacity."

That last phrase is doing work.

What a skill actually is

Strip the branding and a skill is a folder. Inside is a file called SKILL.md with YAML frontmatter — name, description, a few lines of trigger text — plus whatever scripts and reference material a model needs to carry out one task well. The format was built inside Claude Code, where Anthropic says it kept running into the gap between general-purpose intelligence and the procedural knowledge that specific business tasks actually require. In the company's own words, Claude "handled general knowledge work well, but often lacked the procedural knowledge that real business tasks need."

The format is small enough to fit on a cocktail napkin, which may be why 70 agent platforms picked it up before anyone had agreed to govern it. One million skills published over roughly a year of real adoption works out to a new skill about every thirty seconds, around the clock. The Loop will take Anthropic's word on the install count, with the usual caveat that nobody is publishing an audited skills census.

The ten-month playbook, repeated

The arc here is almost suspiciously orderly. In October 2025, Anthropic launched Agent Skills as a Claude Code feature. In December 2025 it published the specification and a reference SDK on agentskills.io, calling it an open standard. In October 2026 it handed governance to the AAIF.

MCP ran the same path about seven months faster: internal convention in mid-2024, open specification shortly after, donation to the AAIF in December 2025. Two protocols, two donations, the same two co-signatories on the receiving end — Block, which Jack Dorsey runs, and OpenAI, whose working relationship with Anthropic is otherwise closer to that of counterparties in a long deposition than to co-maintainers of anything.

The AAIF now holds two Anthropic-origin properties: MCP, donated in December 2025, and Agent Skills, donated this week. It is roughly the open-standards equivalent of the Linux Foundation hosting both Kubernetes and etcd — Anthropic is becoming, in effect, the Google of the agent-protocol stack: the vendor that donates the specifications everyone else builds against, keeps the credit for authoring each, and quietly ensures that any serious agent runtime has to implement its ideas.

Simon Willison, who covered the December 2025 open-standard move, wrote at the time that he "wouldn't be surprised to see this end up" at the AAIF. The prediction was eleven months old before the headline matched it.

The timing

The donation lands between two load-bearing dates in Anthropic's calendar. On October 14, the company began its institutional pre-IPO investor day in San Francisco. Under Section 6(e) of the Securities Act, the public S-1 for a road show starting the week of November 9 needs to be on EDGAR by about October 23. Agent Skills became vendor-neutral the day before that deadline.

The sequencing is not an accident. A specification you still own cannot be a durable moat in a prospectus. A Senate committee, a competitor, or an acquirer can argue that any model-locked format is anticompetitive, and the lawyer talking to the SEC prefers "we donated the standard to a Linux Foundation subsidiary" to "we charge rent on it." The donation closes that door before the filing opens the next one. The IPO narrative becomes: Anthropic is the lab that invents the protocols the industry adopts, and gives them away. Someone will read that as corporate maturity. Someone else will read it as giving away the API lock-in before the equity market can price it in.

Anthropic's own line: "We're committed to keeping Agent Skills open, community-driven, and vendor-neutral." That reads as self-evident until you note who is saying it, where they are saying it from, and that the thing being made vendor-neutral still has their name on every spec page.

Who holds the pen

Governance details worth looking at:

• Lead maintainers: Soria Parra and Murag. Murag was product manager for Agent Skills at Anthropic through its launch; Soria Parra has been one of the public faces of MCP. Both now hold formal governance roles for the standard.
• Co-maintainers from Google and OpenAI, in their individual capacity. The "individual capacity" clause is standard open-source phrasing but preserves plausible-neutrality cover for both sponsors to add or walk back engineering commitment without a corporate signature on either side.
• Membership not required. Anthropic says you do not need to join the AAIF to use or contribute to the standard. That is the single most load-bearing sentence of the governance pitch — it is what keeps 70+ platforms contributing patches rather than negotiating foundation terms.
• Security guidance is still "in proposal." The specification is being made vendor-neutral before anyone has signed off on how to handle untrusted third-party skills. A folder-with-a-markdown-file is a convenient supply-chain surface, and the obvious questions — code signing, provenance, sandboxing of embedded scripts — are formally open items.

What to watch

1. How fast the AAIF publishes the first versioned release. Anthropic references a roadmap of versioned releases, a formal proposal process, and security guidelines. The first Agent Skills release that ships without an Anthropic author line on the top-level spec is the signal that the handover is real and not nominal.
2. Whether the lead maintainers stay Anthropic employees through 2027. The lead-maintainer slot is where the actual steering sits. If both Soria Parra and Murag stay inside Anthropic, the donation is formal. If one moves to a neutral foundation role, it becomes operational.
3. The S-1's language around open standards. Anthropic's filing will have to describe its moat. If MCP and Agent Skills show up in the "competitive advantages" section, the donations are being filed as a moat. If they show up only under "collaboration with the ecosystem," they are being filed as goodwill. Both can be true at once; which the lawyers chose is the signal.
4. Whether OpenAI donates its next protocol. The AAIF's founding contributions were MCP (Anthropic), AGENTS.md (OpenAI), and goose (Block). With Agent Skills, Anthropic has now donated two of its properties against OpenAI's one. If Codex's agent framework or an OpenAI skill equivalent lands in the AAIF next, the three co-founders stay symmetrical. If not, Anthropic becomes the first-among-equals of the foundation it helped build.

The Linux-era pattern was that whoever donated the kernel won, slowly, by making the kernel irresistible to everyone else. The AI-industry version appears to be running at roughly ten-month cadence per protocol. The company moving fastest is the one about to go public.