Anthropic Donates Agent Skills to the AAIF: What Moved and What Did Not

On October 22, 2026, Anthropic shared a short announcement about an important change. The company gave Agent Skills, an open standard it introduced in December 2025, to the Agentic AI Foundation, or AAIF. The AAIF is part of the Linux Foundation and was started by Anthropic, Block, and OpenAI. Exactly four things were handed over: the Agent Skills name, its logo, its GitHub organization, and the agentskills.io website address. Nothing else changed hands. Skills still load the same way, skill folders still use the same format, and skills still work on the same platforms. All of these things worked exactly as they had before the announcement.

This is really a story about who controls a standard and makes decisions about it. People can understand the change in two very different ways. One view is that a format once controlled by a single company now belongs to a wider community. Another view is that only a name and logo changed hands, so the announcement does not matter much. Neither view tells the full story. The announcement gives some unusually clear details: four things were transferred, two people were named as lead maintainers, three future improvements were proposed, and users were told they do not need to join any organization to keep using the standard.

What was actually handed over

The four things Anthropic gave away have different purposes. Understanding those differences helps explain why the transfer matters. The name and logo help decide who can officially call something an Agent Skill. This involves trademark rules, which protect names and symbols. Trademarks can help keep a standard consistent. Without them, different groups might use the same name for things that work differently. The GitHub organization is where people report problems, discuss changes, and see who maintains the project. The agentskills.io domain is the website address where someone can learn about the format, even if they have never installed or used it.

Controlling these four things means controlling the identity and public home of the standard. It does not mean owning the skills people create. A skill is a folder someone makes on their own computer or keeps in their own code repository. Giving away control of the standard does not move or transfer anyone's skill folders. This is an important difference. Anthropic handed over the name and the places where the standard is managed and explained. It did not hand over the files people use in their own projects.

Anthropic explained its goal using three ideas: keeping Agent Skills open, community-driven, and vendor-neutral. The last idea is especially important. Being open means people can access the standard. Being community-driven means people outside one company can help shape it. But a project can claim both things while one company still makes the final decisions. Vendor-neutral means the standard is not controlled to favor one company. Having a foundation oversee the standard can make that promise more believable, because a company cannot prove its own neutrality by itself.

The AAIF already had a history with Anthropic

The AAIF was not created just for this announcement. Anthropic started it with Block and OpenAI last year. Anthropic had also already given another standard to the foundation: the Model Context Protocol. This earlier donation helps explain what is happening with Agent Skills. It shows a pattern. A company creates a format to solve a problem, shares it publicly, sees other people start using it, and then gives control of its name to an organization that includes other companies.

Why do this if the format already works? Because people who use a standard need to trust that its creator will not suddenly change the rules. Imagine building software around a format, only to have the company that created it change that format whenever it wants. For other companies, that would feel more like depending on one company's product than using a shared standard. Giving the name and website to a foundation with several company members makes it harder for one company to reverse course. Changes that were once easy to make privately become harder to make and more visible to everyone.

But this is a way to encourage responsible decisions, not a guarantee. Anthropic says the foundation will give the standard a neutral home and let the community help decide its future. However, the announcement does not explain exactly how community members will make those decisions. The proposed plan includes creating a process for making proposals, but that process has not been explained yet. For now, it is a promise to create a way of making decisions, rather than a finished system.

Who maintains the standard, and why their role matters

Engineers from Anthropic, Google, OpenAI, and other companies help maintain the standard as individuals. David Soria Parra and Mahesh Murag are the Lead Maintainers. The important detail is that these engineers maintain the project as individuals, rather than as official representatives making decisions for their employers.

That difference is not just a technical detail. An engineer acting for an employer may be expected to make decisions that help that company's business. An engineer working through the project's own rules should explain decisions using those rules instead. This does not mean the companies are no longer involved. The engineers still work for those companies, and it would be wrong to say Anthropic has no influence anymore. Agent Skills is not completely independent of Anthropic. What has changed is how maintainers are expected to explain and support their decisions: through the project's rules, rather than simply following company instructions.

It also matters that the maintainers come from several companies that compete with each other. This can make other AI platforms more comfortable using the format. If every maintainer worked for one company, everyone else would have to trust that company to treat them fairly. The more deeply another platform depended on Agent Skills, the bigger that risk would become.

Another important point is that people can keep using Agent Skills or helping improve it without joining the AAIF. Foundation membership is not a special access level, a required payment, or a condition for using the specification. Anyone can still read the specification at agentskills.io. People who want to comment on suggested changes can use the project's Discord, even if they are not foundation members.

What did not change on October 22

Nothing about how Agent Skills works technically changed. This is an important point that could prevent a lot of confusion. Anthropic's announcement changed who oversees the standard. It did not change what the standard does or how people use it.

A skill is still a folder containing a SKILL.md file. That file teaches an AI agent how to do a task or understand a particular subject. The folder can also include scripts and reference materials when needed. The same three-step loading system still applies. First, the skill's name and description, stored in a YAML section at the beginning of the file, are available to the agent when it starts. Second, when a user's request matches that description, the agent reads the main part of SKILL.md. Third, the agent reads extra files mentioned in SKILL.md only when it needs them. Anthropic's ready-made skills for PowerPoint, Excel, Word, and PDF still work exactly as they did before the announcement.

Prices, product packages, and availability also stayed the same. The announcement did not introduce a new pricing level, a new product, or a new endpoint for software to connect to. If you were thinking about doing something new on Thursday that you had not planned to do on Wednesday, this announcement does not give you a reason to change those plans.

This matters because confusing a change in management with a change in technology can waste time and money. Teams might plan unnecessary moves to new systems, rewrite instructions that were already correct, or ask support teams about a change that never happened. For almost everyone reading the announcement, there is nothing they need to do.

The future plans are only proposals

Anthropic has suggested starting with three improvements: releases with version numbers, a formal way to suggest changes, and security advice for people who create and install skills. Each idea addresses a real problem. However, Anthropic has not given dates for any of them. The announcement describes them as possible starting points, not a fixed schedule.

Version numbers may sound boring, but they could be the most useful improvement. When a specification has version numbers, developers can choose which version their software supports, list compatible versions, and check what an AI platform means when it says it supports Agent Skills. Without version numbers, different companies might mean different things depending on which version of the instructions they last read.

A formal proposal process would give people a clear way to suggest and decide on changes. That would make the GitHub organization more than just a place to send messages. But simply saying there will be a process is not enough. The rules matter most. Who is allowed to suggest a change? Who decides whether it is accepted? What happens when a suggestion is rejected? Will everyone be able to see that decision? Anthropic's announcement does not answer these questions. That is why it is important to call the process proposed rather than finished.

Security advice could have the most immediate effect on people's safety, and it may be the first improvement worth watching. Agent Skills can contain instructions and extra resources, including scripts that run code. This means installing a skill is not always just reading a document. It can be more like installing a small software program. Security advice for both skill creators and the people who install skills would help protect both sides. That is a sensible place to begin, although getting everyone to agree on the right rules may be difficult.

Together, these three ideas show how Agent Skills could become a more mature standard. Version numbers would make it easier to depend on. A clear decision process would make changes easier to manage without arguments. Security advice would help people install skills more safely. But these plans also show that the project still has a lot of work ahead.

Why the number of supported platforms matters

According to a summary on Anthropic's news index, more than 70 AI agent platforms have added support for Agent Skills since the standard was published in December 2025. These include ChatGPT, Cursor, GitHub Copilot, and VS Code. This number comes from Anthropic, so it should be treated as a company-reported figure. It counts platforms that have added the format, not individual users. It also does not tell us how much of the format each platform supports or how deeply it is built into their products.

Even if we treat 70 as a minimum number, it helps explain why the change in ownership matters. When a format is used in so many places, the biggest danger is no longer that nobody knows about it. The danger is that different platforms start using it in different ways. If the specification is unclear about something, two platforms might interpret the same instruction differently. People may not notice until they move a skill folder from one AI agent to another and get different results.

A foundation that manages releases with version numbers could help solve this problem. It is also a problem that one company would struggle to solve fairly on its own. When a company writes a standard and also makes a product that uses it, other companies may worry that its decisions will favor its own product. Even if the company tries to be fair, the possible conflict of interest can weaken trust. A foundation with maintainers from several companies that use the standard is in a better position to make decisions people trust. This is one of the strongest reasons for the transfer, even though Anthropic does not say it directly.

This also explains why OpenAI appears in two roles. It helped start the foundation, and it employs some of the engineers who maintain the standard. That overlap is intentional. Agent Skills is supposed to let people use the same skill folder with AI agents made by different companies. Having those companies involved in maintaining the format supports that goal.

What to look for next

The Agent Skills specification is available at agentskills.io, and people can share feedback through the project's Discord. These are the places to watch for future changes. They will be more useful than the original announcement because the announcement explains what Anthropic hopes to do, not the detailed rules for how everything will work.

Three future developments would show that the transfer has made a real difference, rather than just changing the name of the organization in charge. The first would be a published specification with a version number. That would show the plan for versioned releases is becoming real. The second would be a proposal that is publicly rejected. A decision process that has never rejected anything has not yet shown that it can make difficult decisions. The third would be the actual publication of security advice, because that would directly help people who install skills created by others.

Until one of these things happens, the clearest way to describe Agent Skills is this: it is the same technical format, but a different organization now controls its name. Named individuals maintain it, and a new decision-making process has been suggested but not yet explained. This is both a meaningful change and a limited one. That is why people disagree about how important the announcement is.

There is also a reason this matters beyond who manages the standard. OrcaRouter sends AI model requests to providers. It offers access to more than 200 models through one key that works with OpenAI-compatible software, passes through providers' listed prices without adding a markup, and automatically switches between providers when needed. OrcaRouter does not host Anthropic's apps or the places where its skills run. The idea behind Agent Skills is similar to the idea behind a model router. A team should be able to create one skill folder and use it with agents from different companies without changing it. In the same way, a model request should not have to depend on which provider handles it. Shared standards are useful because they make these differences less important. That is also what makes a router useful.

The donation did not improve what Agent Skills can do. Instead, it made it harder for Anthropic to take control of the standard back. For anyone who already has a skill folder stored in a code repository, that is the main benefit. And it matters: the format they depend on is now controlled by an organization that has a reason to keep it working over time, with maintainers from more than one company. The next published version number will help show whether the other promised improvements follow.