Anthropic announced on October 22, 2026 that it is handing over Agent Skills, its open standard for artificial intelligence agents (AI agents), to the Agentic AI Foundation (AAIF), a foundation under the umbrella of the Linux Foundation. More than 70 platforms have already integrated it, including ChatGPT, Cursor, GitHub Copilot and VS Code. The stated aim: to entrust the format to a neutral body that every vendor can use.

What a "skill" is for an AI agent

A skill is a simple folder that teaches an agent a specific task. At a minimum, it contains a SKILL.md file, with a name, a description and instructions written in plain text. The folder can also include scripts, reference documentation and document templates.

Anthropic explains that it designed this format while working on Claude Code, its programming agent. According to the company, the tool handled general tasks well, but it often lacked the procedural know-how of real business tasks: the step-by-step way to get things done.

The technical principle relies on progressive loading, described in the specification published on agentskills.io:

  • Discovery: at startup, the agent reads only the name and description of each skill, roughly 100 tokens (units of text).
  • Activation: when a task matches a description, the agent loads the full instructions, ideally fewer than 5,000 tokens.
  • Execution: the agent follows the instructions and, if needed, runs the attached scripts or files.

This mechanism makes it possible to keep many skills available without cluttering the model's working memory.

Folder elementStatusRole according to the specification
SKILL.md: nameRequired64 characters at most, lowercase letters, digits and hyphens
SKILL.md: descriptionRequired1,024 characters at most; says what the skill does and when to use it
License, compatibility, metadataOptionalTerms of use, required environment, free-form information
allowed-toolsOptional, experimentalPre-authorized tools; support varies from agent to agent
scripts/, references/, assets/OptionalExecutable code, documentation, templates and data

What changes with the foundation

The transfer covers the project's assets, not just its text. Anthropic is handing over the Agent Skills name and logo, the GitHub organization and the agentskills.io domain. The company had published the standard in December 2025, after an initial launch in October 2025.

The AAIF is not a new structure. Anthropic co-founded it in 2025 with Block and OpenAI, under the aegis of the Linux Foundation. It had already contributed the Model Context Protocol (MCP) to it, another standard that connects agents to tools and data.

Maintenance is also changing shape. Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard in an individual capacity. David Soria Parra and Mahesh Murag are its lead maintainers. Using the standard or contributing to it does not require joining the AAIF.

The adoption figures put forward are Anthropic's. The company reports 1 million published skills and nearly 280 million installations. A few weeks after the launch, developers had already written loaders for Codex, Cursor and OpenCode, according to the same source.

The stated commitments and the roadmap

Anthropic says it is "committed to keeping Agent Skills open, community-driven and vendor-neutral". This is a statement of intent: its real scope will depend on the rules the foundation adopts. The same distinction applies to Anthropic and AI regulation, a topic in which 2clics already separated statements from commitments.

For now, the roadmap is only a proposal. It sets out three workstreams:

  • numbered versions of the standard;
  • a formal process for proposing changes;
  • security recommendations for those who write and install skills.

Feedback on specification changes goes through the project's Discord and through GitHub.

What this means for a Swiss company

The main issue is interoperability. The Agent Skills documentation promises to "build a skill once and use it in any compatible agent". A company that describes its internal procedures in this format therefore does not, in principle, lock them in with a single vendor. This matters at a time when vendors are multiplying their offerings, such as OpenAI with its autonomous agents Spaces and Dots.

This portability does have limits, however, stated in the very text of the specification. The allowed-tools field, which pre-authorizes tools, is experimental and its support "may vary" from one agent to another. The languages of the scripts also depend on each agent. A skill that is portable on paper therefore still has to be tested in each tool.

Security deserves the same attention. A skill can contain code that the agent executes. Yet the security recommendations so far appear only in the proposed roadmap. The case of an agent that used a wallet without authorization, documented by OpenAI, is a reminder that an agent equipped with tools can settle an authorization question on its own. Installing a third-party skill therefore means checking its contents, as with any software.

Finally, the format has a practical advantage for a small and medium-sized enterprise (SME): procedures remain text files, versioned and readable by a human. They can be reviewed, audited and corrected by the team, which echoes an observation already made on 2clics: with coding agents, human review becomes decisive.

What we do not know yet

Several points remain open at this stage:

  • the timeline for the first numbered versions and the content of the security recommendations;
  • how decisions will be made between the maintainers and the foundation;
  • the use of Agent Skills in Switzerland: the sources give no figures, as with the adoption of AI agents in Europe, where OpenAI's report did not position Switzerland either.

Finally, the adoption data all come from Anthropic and have not been independently verified. They indicate broad distribution, not actual use within companies.

Sources