Agent Skills Moves to a Foundation: What the Open Standard Means for Your AI Agents

Imagine your team has spent weeks teaching an AI tool how your company reviews a quote, classifies an error message or structures a report. The instructions are neatly written down, with examples and checks. Then the vendor changes its prices, a colleague prefers a different tool, or you want to use two tools side by side. The decisive question is: does what was learned come along, or do you start from scratch? This is exactly where an announcement that Anthropic published on October 22, 2026 comes in: the company is handing the open standard "Agent Skills" over to a neutral foundation.

For managing directors in mid-sized companies, this is not a side issue for developers. Anyone investing in AI agents today is investing, to a considerable extent, in knowledge about their own processes, and that knowledge has to exist in a form an agent can read. Whether that form is tied to a single vendor determines dependency and switching costs. This article sets out what the primary source proves, what it does not say, and which steps you can take this week. A note on availability up front: the standard itself has been published since December 2025 and, according to Anthropic, is used in more than 70 platforms. What is new is the handover of its governance. It is therefore not a vision of the future but concerns tools that exist today.

What Anthropic announced on October 22, 2026

The source is the Anthropic news post "Donating Agent Skills to the Agentic AI Foundation" (anthropic.com/news, October 22, 2026, retrieved by us on October 23, 2026 and read in full). According to this post, "Agent Skills" is an open standard that gives AI agents specialized knowledge and workflows. Anthropic is now handing over the name, the logo, the GitHub organization and the domain agentskills.io to the Agentic AI Foundation, AAIF for short. This foundation is part of the Linux Foundation. According to its own statements, Anthropic co-founded it in 2025 together with Block and OpenAI. The Model Context Protocol (MCP), another open standard that governs how agents talk to tools and data sources, was already handed over to it.

The key facts according to Anthropic at a glance:

• What is being handed over: the name, the logo, the GitHub organization and the domain agentskills.io.
• To whom: the Agentic AI Foundation (AAIF), a project of the Linux Foundation.
• Use and participation: according to the post, anyone who wants to use or help shape the standard does not have to become a member of the AAIF.
• Adoption: more than 70 agent platforms have integrated Agent Skills since its release in December 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code.
• Scale: so far, 1 million skills have been published and almost 280 million installed.
• Maintenance: engineers from Anthropic, Google, OpenAI and other companies maintain the standard as individuals. The post names David Soria Parra and Mahesh Murag as lead maintainers.
• Roadmap proposal: versioned releases, a formal proposal process and a security guideline for people who write and install skills.

On the origins, Anthropic writes that it developed skills while working on Claude Code. In October 2025, thousands of skills were reportedly published on GitHub, and loaders emerged for Codex, Cursor and OpenCode. As its goal, the post names keeping Agent Skills "open, community-driven, and vendor-neutral", that is, open, community-driven and independent of any single vendor.

One qualification belongs here: this is the account of the company that created the standard and is now giving it up. The figures (70 platforms, 1 million skills, almost 280 million installations) come from this post, and we did not count them independently. In addition, the fact that they are so large says nothing about how good individual skills are. A published skill is not automatically a reviewed skill.

What a "skill" is, without the jargon

The post describes skills as a way of giving agents specialized knowledge and workflows. You can think of it like a work instruction for a new employee: how we prepare a quote, in which order we check an invoice, what we pay attention to in a complaint. The difference from the usual chat instruction is that this description exists as a reusable package that the agent can load when the task calls for it. Once written down properly, it is available to every agent that understands the standard.

What a specific skill contains depends on its author. From the source, we only read that it is about knowledge and workflows. The post gives no details on technical file formats in the parts we read, and we are not inventing any here. For the decision in your company, that is not necessary anyway. What matters is the idea behind it: your process knowledge should not disappear into the configuration of a single vendor, but should exist in a form that several tools can read.

What the handover to the foundation means for your company

Until now, the standard rested with the company that invented it. Anyone maintaining their own skills had to trust that this company would keep the standard open. In future, governance rests with a foundation in which, according to the post, engineers from several companies take part. For you as a decision-maker, three practical consequences follow, and we expressly distinguish between what the source says and what we infer from it for your operations.

First: less dependence on one vendor. The source says that more than 70 platforms have integrated the standard, including several that are widely used in German-speaking countries. From this we infer: a well-written skill is no longer an investment in just one tool. We deliberately say "is in principle portable" and not "works identically everywhere". Whether a particular skill runs in one tool exactly as it does in another depends on the tool, and you would have to test that before use.

Second: a negotiating position. If your process knowledge exists in an open format, you can approach vendor selection differently. You then do not ask "Which tool do we choose, and what happens to our instructions?" but "Which tool understands our skills best?" That shifts the balance. At the same time, an open standard does not replace contractual clarification, for example on data protection or on where your data is processed.

Third: a sign of maturity. The fact that a standard sits in a foundation with a roadmap, versioning and a formal proposal process points to a phase of consolidation. We would not derive a guarantee from that, however. The roadmap proposal in the source is a proposal; the source does not describe it as already implemented.

Availability: already usable, but with clear limits on what can be said

Precision is worthwhile here. According to the source, the standard has been published since December 2025 and has since been integrated into more than 70 platforms. New on October 22, 2026 is the handover of governance to the AAIF. The tools named, ChatGPT, Cursor, GitHub Copilot and VS Code, are products that can also be used in Germany, Austria and Switzerland. The post makes no explicit statement about the DACH region; we infer availability from the worldwide use of these tools.

• Status: generally available, open standard since December 2025; what is new is governance by the AAIF.
• Region: usable worldwide via the tools named; the post names no regional restriction.
• DACH: yes, provided the respective tools may be used in your company. You must check this for each tool, for example with regard to data protection.
• Not yet available: the proposed security guideline. It appears in the source as part of a roadmap proposal, not as a finished document.

The last point is the most important one in practice, and we will come back to it shortly.

An open standard reduces dependence on one vendor. But it does not check whether the skill you install does what it promises.

The risk inherent in the number: 1 million skills

One million published skills and almost 280 million installations are impressive figures. But they also mean that there is a huge stock of instructions written by third parties, which an agent executes as soon as it loads them. Unlike a document that you read, a skill is an instruction to software that acts on your behalf and, depending on how it is set up, has access to files, systems or accounts.

A sober consideration follows, which we derive from our work with agents and which does not come from the source: an instruction you have not read should not run unreviewed in an agent that is allowed to carry out real actions. The roadmap proposal in the source apparently acknowledges the issue, since it provides for a security guideline for authors and for people who install skills. The fact that it is named as a proposal shows that until it exists, the review is up to you.

In practice, this means four things:

• Clarify the origin: Who is the skill from? A skill from your own organization is something different from one from an unknown source.
• Read the content: What is the skill supposed to do, which files and systems does it access, what may it change?
• Test in a protected environment first: not in the production system, not with real customer data.
• Document the approval: Who reviewed the skill, when, with what result? Without this trail, in six months you will no longer know why an agent does something.

Concrete examples from mid-sized companies

So that this does not stay abstract, here are three example cases. They are not reports on real projects but thought models that show how the question arises in everyday work.

The mechanical engineering company with its own development department. The developers have written down a workflow for reviewing changes to source code: which tests run, which naming conventions apply, which files are never changed without consultation. Once this workflow is described as a skill, every developer can use it in the tool they prefer, provided it understands the standard. The question for management is not "Which tool?" but "Who maintains the skills, and who approves new ones?"

The service company with recurring reports. A business with 40 employees produces the same analyses every month. The workflow of merging data, flagging anomalies and presenting the result in a fixed structure can be captured as a skill. The advantage: it does not depend on the memory of a single person. The disadvantage, if you are not careful: a faulty step in the skill is then repeated unnoticed every month.

The trade business with digital quote creation. Here, you would use a skill not for code but for workflows such as quote review or appointment scheduling. That is an area in which other agents are responsible. We mention it to make clear that the standard does not concern only programmers. It concerns anyone who wants to capture recurring knowledge in such a way that an agent can use it.

Which agent at Amazing Agents is responsible for this: the Code Operator

On amazing-agents.com, we assign this topic to the Code Operator. The website describes it as: bugs, fixes and pull requests with your approval. It therefore works on the part of your technology where skills and workflows for developer agents typically sit. According to the home page, the same framework applies to all agents: each agent runs within a defined framework with approval steps for critical actions.

What this means concretely for skills, we describe as a procedure, not as a track record: a Code Operator that uses reusable skills would prepare a fix or a change, but would not deploy it into your production code without your approval. That corresponds to what the website promises: pull requests with your approval. New skills would first be reviewed and approved before the agent uses them. This sequence, review first, then approve, then use, is the core.

An agent with limits and an approval step: the difference

The difference between an agent without and one with an approval step can be shown well using skills.

Without an approval step: Someone on the team finds a skill that promises to save effort, installs it and lets the agent work with it. Nobody has read the instructions. The agent executes them as soon as it loads them and can change files or trigger actions that nobody expected. You notice when the result is wrong, and often only then.

With an approval step: A new skill first goes through a review. Someone reads it, checks its origin and access, tests it in a protected environment and then approves it. The agent prepares changes, presents them for review, and a human decides. The work stays fast, and responsibility stays with you. That is an agent with limits and an approval step, and that is exactly what distinguishes it from mere automation that simply runs.

An open standard helps with the first question, namely switchability. With the second, security, it does not help, at least not on its own. The source proposes a security guideline for this. Until it exists, discipline in your own company decides.

What is still to come: what we do not know

An assessment like this honestly also includes what remains open.

• Timeline of the security guideline: The source names it as a proposal, but gives no date.
• Quality of the published skills: The figure of 1 million says nothing about their quality or security.
• Behavior of individual tools: Whether a skill works the same in every tool depends on the tool. The source does not say, and we have not tested it.
• Legal and data protection questions: Whether and how you may use a particular skill with customer data is not settled by any standard, but by your data protection adviser.
• Maintenance effort: Anyone who introduces skills has to maintain them. How much effort that means day to day is not stated in the sources.

What you can do this week: three steps

Because the standard is already usable, these are not mere preparatory steps but things you can set in motion without any procurement.

1. Collect process knowledge: Have each team write down two to three recurring workflows where knowledge so far exists only in people's heads. Include: Who carries out the workflow, what steps are there, where are decisions made? This is the raw draft of a skill and has value even without AI.
2. Set a rule for third-party skills: Decide in writing that no skill from an unknown source may run in an agent that can change files, send messages or address systems unless someone has read and approved it. Name a person who decides this.
3. Designate a test environment: Clarify where new skills can be tried out without consequences, without real customer data and without access to production systems. Without such a place, testing in practice happens directly in live operations.

Conclusion

The handover of "Agent Skills" to the Agentic AI Foundation is good news for everyone who wants to invest process knowledge in agents without being chained to one vendor. The source proves: the standard has been published since December 2025, is used in more than 70 platforms according to Anthropic, comprises 1 million published skills and will in future rest with a vendor-neutral foundation. It does not prove that every skill is safe, and it names the security guideline as a proposal, not as something that exists.

For you, this means: use the switchability you have gained, but keep control. Write down your knowledge in such a way that an agent can read it. Review every skill before it runs. And for everything that changes something, have a human approve. An agent with limits and an approval step is not a brake but the precondition for being able to hand it more work in the first place.

Sources: Anthropic, "Donating Agent Skills to the Agentic AI Foundation", anthropic.com/news, October 22, 2026 (retrieved on October 23, 2026). Assignment to the agents: amazing-agents.com, home page (Code Operator; approval steps for critical actions).