Anthropic announced on 22 October 2026 that it is handing Agent Skills — the open standard for giving an agent specific knowledge and procedures — to the Agentic AI Foundation (AAIF), the body hosted by the Linux Foundation. The handover is not symbolic: it covers the name, the logo, the GitHub organisation and the agentskills.io domain. According to the company, more than seventy agent platforms have adopted the format since it was published as a standard in December 2025, among them ChatGPT, Cursor, GitHub Copilot and VS Code.
What was announced
The thing being donated is a deliberately thin format. A skill is, in essence, a folder: inside sits a file called SKILL.md that tells the agent how a task is done or how a domain works, next to whatever scripts or reference documents that task needs. No server, no transport protocol, no mandatory central registry. That simplicity is, Anthropic argues, why so many products could adopt it without negotiating with anyone.
The figures the company offers to justify the move are two: one million skills published and close to two hundred and eighty million cumulative installs. Neither comes with a breakdown or a counting method.
The AAIF becomes the standard's home. Anthropic stresses that joining the foundation is not required to keep using the format or to propose changes: both use and contribution stay open. Maintenance falls to engineers from Anthropic, Google, OpenAI and other companies, but acting as individuals rather than on behalf of their employers. David Soria Parra and Mahesh Murag stay on as the project's lead maintainers.
The proposed starting roadmap has three strands: shipping numbered releases of the standard, setting up a formal route for proposing and accepting changes, and writing security guidance aimed at both the people who author skills and the people who install them. Discussion of specification changes runs through the project's Discord, with the repository still the place where detail gets argued out. The specification is published at agentskills.io.
Where it comes from
Agent Skills did not start life as a standard but as a patch for a practical problem. Anthropic says it developed the idea while working on Claude Code: the model handled generic knowledge work well enough, but lacked the procedural knowledge that real business tasks demand — the how we do this here, which sits in no public corpus.
The initial launch was in October 2025. Within weeks, the company says, thousands of skills were up on GitHub and third parties had written loaders for Codex, Cursor and OpenCode. In other words, the ecosystem ran ahead of the specification. Anthropic responded by moving the format out of its own house in December 2025: a separate GitHub organisation, a separate domain and open-standard terms so any agent could implement it.
The AAIF itself is neither new nor unfamiliar territory for Anthropic. The foundation was set up last year with Block and OpenAI among its founders, and it is where the Model Context Protocol (MCP) was already deposited. Today's move repeats the play with the company's second standard.
What it means
For anyone building or maintaining software with agents, the headline matters less than the shift in risk. Until now, any team with an internal catalogue of skills depended on a private company not changing the format to suit its own product plans. With the trademark, the repository and the domain in the hands of a neutral foundation, that risk drops — it does not vanish, but it drops.
The second practical consequence is versioning. Until today the standard was whatever sat on the repository's main branch: a living format with no number attached. If the plan holds, there will be tagged releases to validate against, which makes it possible to write in internal documentation that "our skills target version X" and to spot breakage on upgrade.
The third is security, and it is the one that should move anyone running this in production. A skill is code and text entering the context of an agent that holds permissions. Installing one from a third party is, in attack-surface terms, close to installing a dependency: somebody writes instructions your agent will follow. That the standard now acknowledges the need for security guidance is itself confirmation that the gap exists and that today each team patches it however it can.
What the announcement does not say
There is no effective date for the transfer, and no detail on how the project will be governed inside the foundation beyond the lead maintainer roles. Nor is there any word on what happens to the skills already out there: no official registry is mentioned, no signing or verification mechanism, and nothing about the catalogues currently living in each platform's private repositories.
The numbers are claims from an interested party. Anthropic does not clarify what counts as a published skill — whether duplicates or forks are filtered out — or what counts as an install. The list of integrated products also comes from the company.
Finally, the classic friction point of these moves stays open: what happens when a vendor's own agent needs an extension to the format before the formal proposal process has approved it. The announcement promises transparent decisions and community weight, as with MCP, but does not describe what gives when a product calendar and a standards calendar disagree.
The original announcement is published on Anthropic's site.