Anthropic is transferring Agent Skills to the Agentic AI Foundation, moving the reusable workflow standard into a neutral governance structure. The October 22 announcement accompanies a wider expansion of the foundation’s projects covering how AI agents acquire capabilities, connect to models and operate.
TL;DR: Agent Skills is joining AAIF alongside three other projects. The change concerns ownership and governance of a shared format; it does not require users to replace their existing agents.
A common format gets a broader home
Anthropic says its donation includes the Agent Skills name and logo, its GitHub organisation and the agentskills.io domain. Participation will remain possible without joining the foundation. David Soria Parra and Mahesh Murag are named as lead maintainers.
The company reports that more than 70 platforms support the format, with one million skills published and nearly 280 million installations. Those figures describe reported ecosystem activity; they should not be read as a count of unique users or independently verified business deployments.
The Agent Skills documentation explains the format in concrete terms. A skill is a folder containing a SKILL.md file with a name, description and instructions. It can also contain scripts, reference documents and other resources needed to complete a task.
Agents initially load a skill’s identifying information, then retrieve fuller instructions when a task calls for them. This approach lets teams retain procedures in reusable files instead of restating every requirement whenever they ask an agent to work.
A useful distinction follows: a skill packages instructions and context, rather than providing a guarantee of correct execution. The quality of the procedure, the information available and the agent’s permissions still determine whether a particular workflow is appropriate.
Four contributions address different parts of the stack
The foundation’s announcement also accepts OpenAI’s Open Responses, Hugging Face’s OpenEnv and Alibaba’s OpenSandbox. Together with Agent Skills, the additions take AAIF’s hosted portfolio to 10 projects, compared with three at its December 2025 launch.
The contributions cover different problems: portable expertise, interfaces across model providers, environments for training and evaluation, and places to execute agents. AAIF announced them at AGNTCon and MCPCon North America in San Jose.
For developers, neutral governance creates a forum in which competing vendors and users can discuss changes to shared infrastructure. It does not make every implementation interchangeable automatically. Teams still have to check how their chosen products support a specification and handle the resources a workflow needs.
Anthropic’s proposed next steps include versioned releases, a formal proposal process and security guidance. Those are roadmap items. Their implementation will help determine how manageable the standard becomes for organisations that want to maintain and review workflows over time.
Written by the News Desk and edited by Samuel Rae.
Leave a Reply