Anthropic announced on October 22, 2026 that it was donating Agent Skills to the Agentic AI Foundation (AAIF). The foundation confirmed Agent Skills as a new project, giving the format for reusable agent instructions a neutral governance home. For builders, the immediate change is stewardship. The announcements do not establish that a skill will behave identically in every agent product.
What moves to AAIF
Anthropic says its donation includes the Agent Skills name and logo, GitHub organization, and agentskills.io domain. It says engineers from Anthropic, Google, OpenAI, and other companies will maintain the project as individuals, with David Soria Parra and Mahesh Murag as lead maintainers.
Anthropic has proposed a roadmap for versioned releases, a formal proposal process, and guidance for people who write and install skills. Those are plans, not completed governance or security controls. The announcement does not introduce a new specification version or a compatibility guarantee.
What the format specifies
The published Agent Skills specification defines a skill as a directory containing a SKILL.md file. That file needs YAML frontmatter with a name and description, followed by instructions. Scripts, references, templates, and other assets are optional. The specification also lists an experimental allowed-tools field and warns that support may differ between clients.
This is a format for packaging procedural knowledge. It does not, by itself, define a shared permission model, prove that bundled code is safe, or guarantee that two products will execute the same instructions the same way.
Where portability stops
The project’s client implementation guide describes discovery, loading, and activation, but leaves implementation choices to each client. A local agent may scan directories on a user’s machine; a cloud agent needs another way to obtain the same files. Clients can also differ in where they look for skills and how they give an agent access to bundled resources.
A team moving a skill between products should therefore check the installation path, whether the client discovers and activates it, which resources it can read or run, and what permissions govern its actions. A shared folder structure is useful, but it is not a substitute for testing those behaviors in each target client.
What builders should do now
A working skill does not need to be rewritten solely because its standard has a new steward. Keep track of the skill revision and the client versions you tested. Review instructions and bundled scripts from outside your organization before granting them access to files, tools, or services.
The proposed security guidance may help future authors and installers. Today, teams still need their own trust checks, permission boundaries, and tests for the particular clients in which they run skills.
