Anthropic hands Agent Skills to the AAIF: the skills standard for agents moves to neutral governance

By DiarioBitcoin · October 22, 2026.

Anthropic has announced the donation of Agent Skills to the Agentic AI Foundation (AAIF), the foundation that operates under the Linux Foundation and that Anthropic, Block and OpenAI co-founded in 2025. According to the company, the transfer includes the name, the logo, the GitHub organization and the agentskills.io domain.

A "skill" is a folder containing a SKILL.md file and, where needed, scripts or reference materials. It lets an AI agent load specialized procedures when it needs them. It is not a new model: it is a package of instructions and resources for a specific task.

The adoption figures come from Anthropic and have not been independently verified. The company says more than 70 platforms integrate the standard, including ChatGPT, Cursor, GitHub Copilot and VS Code. It also reports one million published skills and nearly 280 million installs, without specifying how many distinct users or agents are behind that figure.

Timeline according to Anthropic: skills originated in Claude Code, whose initial launch it dates to October 2025. Developers published thousands on GitHub within a few weeks and built loaders for Codex, Cursor and OpenCode. In December 2025 it was released as an open standard.

Maintenance will fall to engineers from Anthropic, Google, OpenAI and other companies, who will participate in an individual capacity rather than as representatives of their companies. David Soria Parra and Mahesh Murag will be the Lead Maintainers. Using or contributing to the standard does not require joining the AAIF.

The proposed roadmap includes numbered versions, a formal proposal process and security guidance for those who write or install skills. These are initial proposals, not approved rules. The announcement does not detail controls, technical requirements or a timeline.

**Our take.** This is another piece of agent "plumbing" that Anthropic is handing to the AAIF, following MCP (Model Context Protocol, the protocol for connecting agents to tools and data). It fits a thesis we hold: the agent wars will be decided less by the smartest model and more by who controls the integration standards. Donating them means giving up formal control, but also ensuring the format is the one everyone uses. It is a smart move and, at the same time, genuinely good for the ecosystem: a portable format keeps each vendor from locking its users into its own skills catalog.

The sensitive point is security. A skill can include executable scripts, so installing one from a third party is a lot like installing a software package of unknown origin. With a million skills in circulation, the supply-chain risk is real and near-term: it is natural ground for automated fraud and prompt injection. That security is on the roadmap is a good sign, but today it is an intention without concrete controls. Signatures, verifiable provenance and scoped permissions should be required before volume turns the problem into an incident.

We should also be cautious about the "neutrality" framing. The lead maintainers are still from Anthropic, and real governance will be measured by whether other players manage to change the specification against its creator's interests. Only the first formal proposals will show that.

In the long term, the outlook is favorable. Open standards for reusable capabilities lower the cost of building competent agents in any domain, including medicine and research, and spread that power among many instead of concentrating it in a few. It is a modest brick, but one of those that make abundance possible, as long as security is built alongside scale.