Anthropic transferred control of Agent Skills, its open format for packaging agent instructions, to the Linux Foundation's Agentic AI Foundation on October 22.

The format already runs inside ChatGPT, Cursor, GitHub Copilot and VS Code. What changed hands is governance. The name, the logo, the GitHub organization and the agentskills.io domain now belong to a vendor-neutral foundation instead of the company that built them.

What actually moved

Anthropic shipped skills inside Claude Code in October 2025. Developers published thousands of them within weeks and wrote loaders for Codex, Cursor and OpenCode, so the company published the format as a standalone standard that December. It now says more than 70 agent platforms have integrated it, 1 million skills have been published, and those skills have been installed nearly 280 million times.

The donation is narrower than those figures suggest. Anthropic gave up the trademark and the infrastructure, not the reference implementation, and its post states that anyone can keep using or contributing to the standard without joining the foundation.

"Without the right context, even the most intelligent agents fall short," said David Soria Parra, co-creator and lead maintainer of MCP at Anthropic.

Soria Parra and Mahesh Murag are the lead maintainers. Engineers from Anthropic, Google, OpenAI and other companies maintain the project as individuals. That structure keeps one vendor from owning the spec, and it also makes maintainer time the scarce resource.

Simon Willison noted on October 22 that he had predicted the transfer in December 2025, when Anthropic first called skills an open standard.

Why a folder travels further than an API

A skill is a directory containing a SKILL.md file. The frontmatter needs two fields: a name of up to 64 lowercase characters and hyphens, and a description of up to 1,024 characters that says what the skill does and when to use it. Optional fields cover the license, environment requirements, metadata and a pre-approved tool list that the spec still marks as experimental.

The design depends on lazy loading. Agents read only the name and description at startup, roughly 100 tokens per skill, then pull the full body when a task matches. The spec asks authors to keep that body under 5,000 tokens and the file under 500 lines, with scripts and reference documents loaded only when the agent reaches for them.

That token budget explains the adoption. Teams can ship dozens of procedures without paying for all of them in every prompt, and the same folder runs in any client that implements the format. A reference validator called skills-ref checks the frontmatter before a skill ships.

Adoption numbers hide a quality question

A million published skills say little about whether they help. A study submitted to arXiv on October 4 examined 3,159 skills and 2,608 revision pairs to measure what a revision does to an agent.

The authors found that 55% of revisions add, remove or rewrite a rule. Across 16 open-weight models, adding one checkable rule raised compliance by 0.41 on average. Across four agents working in a sandbox, the rate at which the agent took the required action rose 0.23, and blind judges recorded a 0.10 rise in final correctness.

Most of the gain came from rules that named a command or a path the previous version left out. When the harness loaded a skill body only on demand, as real tools do, agents kept about half of the action gain and open models about 38%, while loading a body added 50% to the tokens of an episode.

The format works, but vague instructions change nothing and bloated files waste context. Anthropic's install count measures distribution, not usefulness.

A roadmap is not a specification

Anthropic proposed a first roadmap: versioned releases, a formal proposal process and security guidance for people who write and install skills. Neither the company post nor the foundation release gives a date, and no versioned specification exists yet.

The security item matters because a skill can bundle executable scripts. Installing one from a stranger is closer to installing a package than to reading documentation. The experimental allowed-tools field does not settle that on its own, since implementations handle the field differently.

The foundation now hosts 10 projects, up from three at its December 2025 launch. The four announced on October 22 stretch across the agent stack: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. MCP, A2A, AGENTS.md, goose, agentgateway and Agent Router were already there.

"A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10," said Manik Surtani, chief technology officer and co-founder of the foundation.

What teams should do now

• Record which spec revision each skill targets. Versioned releases are planned, and no numbering scheme exists yet.
• Read the scripts in any third-party skill before it runs on a machine that holds credentials.
• Test one skill on two agents. Portability stays a claim until someone checks it.
• Watch the proposal process once it opens, because field changes will arrive through it.

Existing skills keep working, and the move costs users nothing. The open question is what the first security guidance will require, and whether it lands before the install count doubles again.

Sources

1. Anthropic, "Donating Agent Skills to the Agentic AI Foundation," October 22, 2026
2. Linux Foundation, "Agentic AI Foundation Announces Four New Projects," October 22, 2026
3. Agent Skills specification, agentskills.io
4. Wang et al., "Agent Skill Evolution: How Revisions Affect Coding Agents," arXiv:2610.04832
5. Agentic AI Foundation, hosted projects index
6. Simon Willison, "Donating Agent Skills to the Agentic AI Foundation," October 22, 2026