Anthropic gave Agent Skills to the Agentic AI Foundation

On October 22, Anthropic donated Agent Skills to the Agentic AI Foundation. Anthropic says the foundation sits under the Linux Foundation, and that it co-founded the foundation with Block and OpenAI. A skill is a folder that teaches an agent one task or one domain. Which tasks belong there is covered in What to hand to AI first.

What moved

Anthropic published Agent Skills as an open standard in December 2025. On October 22 it gave that standard to the foundation: the name, the logo, the GitHub organization, and the agentskills.io domain. Anthropic says people can keep using the standard, and can contribute to it, without becoming members of the foundation.

The foundation’s release the same day says it is taking in four projects. Agent Skills comes from Anthropic. Open Responses comes from OpenAI. OpenEnv comes from Hugging Face. OpenSandbox comes from Alibaba. The foundation says those additions bring it to 10 hosted projects, up from the three it launched with in December 2025.

Anthropic says engineers from Anthropic, Google, OpenAI, and other companies will maintain the skills standard as individuals. It names David Soria Parra and Mahesh Murag as lead maintainers. It has proposed versioned releases, a formal proposal process, and security guidance for people who write and install skills. In this announcement, that guidance is a proposal.

What a skill contains

Anthropic defines a skill as a folder. A file named SKILL.md teaches one task or one domain. Scripts and reference files can sit beside it.

Anthropic says it built skills while working on Claude Code. Claude Code handled general knowledge work, and it often lacked the procedural knowledge a real business task needs. In Anthropic’s company-reported account, developers published thousands of skills on GitHub within weeks of an October 2025 launch, and wrote loaders for Codex, Cursor, and OpenCode. Anthropic then published the format so any agent could use it.

Two further figures are company-reported. Anthropic says more than 70 agent platforms have integrated the format since the December 2025 standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code. It says 1 million skills have been published and nearly 280 million have been installed.

Those figures count skills published, platforms that integrated the format, and installs. They are Anthropic’s. The foundation describes a skill as instructions, scripts, and resources an agent can load when needed, so know-how can move across products. Whether a given folder moves is a check on one job.

The other three, in short

Open Responses, from OpenAI, is a specification for talking to more than one model provider. The foundation says it is based on OpenAI’s Responses API: a shared schema, consistent streaming events, and extensible tooling, so a request can be described once and run across providers with minimal translation.

OpenEnv, from Hugging Face, is a framework for the reinforcement-learning environments used to train and grade agents. OpenSandbox, from Alibaba, is a sandbox platform. The foundation says OpenSandbox includes a protocol for lifecycle and execution, plus Docker and Kubernetes runtimes.

On October 22 the foundation took in the format for the written job, the spec for the model call, the training environment, and the sandbox, under governance it calls vendor-neutral. The folder is the piece an owner can write this week.

What the folder still needs

Anthropic traces skills to a gap in procedural knowledge. That procedure still has to come from the business. For one job the team already repeats, the folder has to hold three things:

• The trigger, in one sentence a teammate would recognize.
• The steps, what done looks like, what the agent may open, and when a person takes the work back.
• One reference the model would otherwise invent: a price, a status name, or a file path.

The guide holds the rule for which job: it happens often, follows clear rules, is easy to check, and can be reviewed before it reaches a customer.

A skill someone else wrote is a folder an agent may load, scripts included. The open question is who published it and what the scripts can reach.

One folder, run twice

1. Name the folder after the job.
2. In SKILL.md, write the trigger, the steps in order, and a description of done a person can check without rerunning the agent.
3. Add one reference file. Add a script only when someone on the team already runs that script by hand.
4. Point the current agent product at the folder and save the output.
5. If a second product from the set Anthropic named is already in use (ChatGPT, Cursor, GitHub Copilot, or VS Code), point it at the same folder. Record the first failure, or record that it ran.
6. With only one product, stop after step 4. The job is in the format Anthropic published so any agent could use it, and that the foundation describes as portable across agent products and platforms. Whether this folder transfers is still untested.

A second run that succeeds covers that one job on those products. It is a note about this folder.

What to do this week

• Write that folder for one job the team already repeats.
• Save the first run. Where a second product is already in use, save what differed.
• List any installed skill nobody on the team wrote, and note the publisher.
• Keep the current model. The announcement gave the skills standard to the foundation.

A folder is easier to carry than a chat someone has to paste again. A consult block includes picking the one or two workflows where AI saves the most time or money, and working on them with you.