AI news · Source-checked implications

Agent Skills moves to AAIF: what transfers between tools?

The standard moves to a neutral foundation. The file format is portable; runtime requirements, permissions and successful execution still need checking.

Edited by Clinton FeyisitanPublished Announced 2026-10-22

Review scope: Official announcement and linked technical documentation read 28 October 2026. No installation, token creation, paid request or native workflow test. Dates describe the announcement and our source read separately.

The change is governance

Anthropic announced on 22 October that it is donating the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation. The contribution includes the name and logo, GitHub organisation and domain. Anthropic says using or contributing to the standard does not require joining the foundation.

Versioned releases, a formal proposal process and security guidance are proposed roadmap items in that announcement. They should not be described as already delivered safeguards.

What a skill actually carries

The format packages instructions in a SKILL.md file, optionally accompanied by scripts, references and assets. The overview describes staged loading: an agent discovers names and descriptions, reads full instructions when relevant, then uses the supporting resources as needed.

The specification requires a name and description. Compatibility and licence fields are optional. The allowed-tools field is explicitly experimental. A shared folder format therefore does not establish identical permission behaviour in every client.

What to check before moving a skill

Our practical implication is to separate format compatibility from workflow success. Inspect the intended client, required software, network access, bundled scripts and licence before reusing a skill. Check how that client requests approval and which actions it can execute. A missing compatibility field is not proof that the skill works everywhere.

Run a small task with non-sensitive fixtures before relying on the result. Compare the output against your acceptance criteria and record the exact client and skill version. These are proposed validation steps; Fewertools has not executed a cross-client test for this report.

For related boundaries, see Claude Code permissions and sandboxing, Cursor Agent SDK deployment roles and coding tools. The donation does not establish a new model price, free allowance or comparative winner.

Sources and review scope

This is a documentation report. It does not establish performance, security or reliability in your environment.