Agent Skills, the format Anthropic created so AI agents can load packaged know-how on demand, no longer belongs to Anthropic. On October 22 the company said it had placed the standard with the Agentic AI Foundation (AAIF), the body under the Linux Foundation that it set up last year with Block and OpenAI and that already looks after the Model Context Protocol. The move was announced at AGNTCon + MCPCon North America in San Jose, where the foundation also accepted three other projects.
A skill is simple by design: a folder holding a SKILL.md file that describes a task or domain, plus any scripts or reference files the agent needs to carry it out. Anthropic says the idea grew out of work on Claude Code, which handled general tasks well but lacked the procedural knowledge many business jobs require. After skills appeared in October 2025, developers wrote loaders for Codex, Cursor and OpenCode, and Anthropic published the format as an open standard in December 2025.
Adoption figures in the announcement are Anthropic's own. The company counts more than 70 agent platforms that support skills, among them ChatGPT, Cursor, GitHub Copilot and VS Code, and reports roughly 1 million published skills and close to 280 million installs. Those numbers have not been independently audited.
What changes hands is concrete. The foundation receives the brand itself (name plus logo), the project's GitHub organization and its agentskills.io web domain. People can keep using or contributing to the standard without joining AAIF. Maintenance falls to individual engineers drawn from Anthropic, Google, OpenAI and elsewhere, acting in a personal capacity, with Anthropic's David Soria Parra and Mahesh Murag named lead maintainers. The first proposed roadmap items are versioned releases, a formal process for proposing changes, and security guidance for people who write or install skills.
Skills were not the only arrival. According to the Linux Foundation, AAIF also took in Open Responses, an OpenAI-contributed specification based on the Responses API that aims to let developers describe requests once and run them against several model providers; OpenEnv from Hugging Face, a framework and hub for packaging reinforcement learning environments as containerized services; and OpenSandbox from Alibaba, aimed at secure execution for agents. The foundation says it now hosts 10 projects, up from three at launch in December 2025, alongside MCP, the A2A protocol, AGENTS.md and goose.
For builders, the practical effect is that a growing share of the plumbing agents depend on, from tool connections to instructions to sandboxes, now sits under one neutral governance roof rather than with individual vendors. That may make it easier for rival platforms to commit to the same formats.
Open questions remain. Neutral governance does not by itself settle how skills are vetted, and the security guidance Anthropic mentions is still a proposal, even though skills can bundle executable scripts. How quickly the foundation turns its four new arrivals into versioned, broadly implemented specifications will show whether the consolidation delivers more than a change of ownership.