Anthropic hands Agent Skills over to the Agentic AI Foundation

What it's about

On October 22, 2026, Anthropic announced it is donating the open Agent Skills standard to the Agentic AI Foundation (AAIF), a foundation under the Linux Foundation. Anthropic says it wants to keep Skills "open, community-driven and vendor-neutral." Anthropic co-founded the AAIF in December 2025 with Block and OpenAI, and had already contributed the Model Context Protocol (MCP) to it at the time. Blogger Simon Willison noted that he had predicted this move in December, when Anthropic first called Skills an open standard.

What Agent Skills are

Agent Skills is an open format for giving AI agents knowledge and workflows for specific tasks. Anthropic introduced it as a Claude feature in October 2025 and released it as an open standard in December 2025. A skill is a folder built around a SKILL.md file containing the instructions. Scripts, reference documents and templates can be added, like this:

my-skill/
├── SKILL.md
├── scripts/
├── references/
└── assets/

SKILL.md contains the name, description and execution steps, among other things. One advantage: agents load only the skills relevant to a task. That lets you use many skills without needlessly filling the context window (the amount of text a model can process at once).

Adoption and next steps

According to Anthropic, more than 70 agent platforms have integrated Skills since December 2025, including ChatGPT, Cursor, GitHub Copilot and Visual Studio Code. With the donation, the standard is meant to be developed jointly under neutral governance. According to a summary on note.com, the next steps are versioning of the specification, a formal proposal process and security guidelines. David Soria Parra, co-creator and lead maintainer of MCP at Anthropic, said the company wants to keep investing in open agent tooling.

The AAIF grows to ten projects

Alongside Skills, the AAIF announced three more new projects. It now has ten, up from the three launch projects of December 2025:

• Open Responses defines a common schema, unified streaming events and extensible tools, so developers can describe requests once and use them with different model providers with little translation effort.
• OpenEnv exposes RL environments as a containerized service that offers a familiar agent interface (MCP) and the Gym training protocol. According to Hugging Face, participants include Microsoft, Meta-PyTorch, Modal and Prime Intellect.
• OpenSandbox serves coding and GUI agents, agent evaluation, code execution and RL training, and includes a sandbox protocol for lifecycle and execution that can be extended with custom runtimes.

Existing projects include the Agent2Agent protocol (A2A), Agent Router, agentgateway, AGENTS.md, goose and MCP. AAIF CTO Manik Surtani said the projects cover different layers, from what agents can do to where they run. The news was announced at AGNTCon + MCPCon North America (October 22–23, San Jose).

How the standards fit together

The note.com summary describes the three best-known standards as complementary rather than competing: if you ask an agent to analyze internal codebases and write a report, it reads project rules via AGENTS.md, pulls in analysis steps via Agent Skills and connects to external services via MCP.

Our take

The donation is a logical step: a standard already supported by competitors such as OpenAI, Microsoft (GitHub Copilot) and Cursor looks more credible when no single company controls it. For developers, this mainly means predictability: skills you write once should work across products. The source also notes a caveat: because of differences in execution environments and supported features, not every skill runs unchanged everywhere. On the critical side, the figure of more than 70 platforms comes from the vendor, and the security guidelines are only planned. That matters, especially for skills that include scripts. Whether neutral governance really opens up development will show in the proposal process and in contributions from outside the founding companies.