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

On 22 October 2026, Anthropic published a short post with a long reach. The company donated Agent Skills, the open standard it published in December 2025, to the Agentic AI Foundation — the AAIF — a body under the Linux Foundation that Anthropic co-founded with Block and OpenAI. Four named assets changed hands: the Agent Skills name, the logo, the GitHub organisation, and the agentskills.io domain. Nothing else did. The way a skill loads, the format of a skill folder, the surfaces a skill already runs on — all of it was identical the day after the post to the day before it.

That makes this a governance story, and governance stories get inflated in both directions. The generous reading is that a vendor-controlled format has become a community standard. The cynical reading is that a logo moved and everyone should go home. Neither survives contact with the detail, because the detail is unusually specific for an announcement of this kind: four assets, two named lead maintainers, three proposed roadmap items, and an explicit line saying you do not have to join anything to keep using the standard.

What was actually transferred

The four assets are not equivalent, and the differences are where the meaning sits. The name and the logo decide who may call something an Agent Skill. That is a trademark question, and trademarks are the mechanism by which a standard stays a standard instead of splitting into dialects that each claim the label while behaving differently. The GitHub organisation is the working surface — where issues are filed, where changes are discussed, and where the identity of the people who maintain the project is published. The agentskills.io domain is the reading surface, the address someone visits when they have never installed anything and want to know what the format is.

Custody of those assets is what owning a standard means. It is not custody of anyone's skill. A skill is a folder that somebody wrote on their own machine or in their own repository, and donating the standard does not move a single one of them. This is the first place the announcement is easy to overread: what Anthropic handed over was the identifier and the address, not the contents of anyone's working directory.

Anthropic described the motive in one phrase — keeping Agent Skills open, community-driven, and vendor-neutral. It is worth noticing the third word, because the first two describe a process and the third describes an outcome. Open and community-driven are things a project can claim while one company still holds the pen. Vendor-neutral is the claim a foundation exists to make credible, and it is the one that cannot be made by the vendor alone.

The destination had a history with Anthropic already

The AAIF is not a body convened for this occasion. Anthropic co-founded it with Block and OpenAI last year, and it had already received one Anthropic donation: the Model Context Protocol. That precedent matters for how to read the Agent Skills transfer, because it establishes a pattern rather than a gesture. A company builds a format to solve its own problem, publishes it, watches other people build on it, and then hands the name to a body with other members in it.

Why bother, if the format already works? Because a standard's value to everybody who is not the publisher is inversely related to how cheaply the publisher can change it. A format one vendor can redefine on a Tuesday is, from the position of a team building on it, a vendor feature with extra paperwork. Moving the name and the domain into a foundation with several corporate members is a commitment device: it makes the cheap reversal expensive and the expensive reversal public.

It is a commitment device, not a guarantee. The announcement says the foundation will give the standard a neutral home and ensure the community has a say in where it goes next, but it does not describe what that say looks like. A roadmap that proposes a proposal process is a promise about a promise, and the honest way to report it is as exactly that.

Who maintains it, and the phrase that carries the weight

Engineers from Anthropic, Google, OpenAI and other companies maintain the standard as individuals, with David Soria Parra and Mahesh Murag as Lead Maintainers. Two words in that sentence do most of the work: as individuals.

It would be easy to read that as a formality. It is not. It is the difference between a maintainer acting as an employee of one company with a commercial interest in the outcome, and a maintainer acting under the project's own process. It does not mean the companies have walked away — the maintainers still work for them, and pretending otherwise would be precisely the overstatement to avoid here. The standard is not now independent of Anthropic. What changed is the channel through which a maintainer's decision has to be justified: project process rather than employer instruction.

The composition of the maintainer set is the other half of the point. A format maintained by engineers from several competing agent platforms is one those platforms can afford to build on. If the maintainers all sat inside a single vendor, every other adopter would be betting on that vendor's goodwill, and the bet would get more expensive as their integration got deeper.

One more line deserves attention: you can continue to use the standard or contribute to it without joining the AAIF. That closes the most common misreading in advance. Joining is not a membership tier, a fee, or a gate on the specification. The specification stays readable at agentskills.io, and the feedback channel for proposed specification changes is the project's Discord, open to people who have not joined the foundation.

What did not change on 22 October

Nothing technical. That is the sentence that would have saved a number of commentary threads, and it is the one Anthropic's post does not shout. The announcement moved governance. It did not move behaviour.

A skill is still a folder containing a SKILL.md file that teaches an agent a task or a domain, plus any scripts or reference material the skill needs. The three-level loading order is unchanged: the name and description from the YAML frontmatter sit in the agent's context from startup, the body of SKILL.md is read when a description matches a request, and additional files referenced from that body are read only when something points at them. The pre-built document skills Anthropic ships for PowerPoint, Excel, Word and PDF run on exactly the same mechanism they ran on the day before.

Nor did anything change in price, packaging or availability. There is no new tier, no new product and no new endpoint in the post, because there was not one. If you were planning on Thursday to do something you had not planned on Wednesday, this announcement is not the reason.

That is worth saying plainly, because reading a governance change as a capability change has a cost. Teams schedule migrations nobody asked for, rewrite documentation that was already correct, and open support threads about a transition that did not happen. The correct action item for almost every reader of this post is none.

The roadmap is proposed, and that is the honest word for it

Anthropic has proposed starting with a roadmap covering three items: versioned releases, a formal proposal process, and security guidance for people who write and install skills. Each is a concrete answer to a real problem. None of them has a date, and the post presents all three as a starting point rather than a schedule.

Versioned releases are the least glamorous item and probably the most useful. A specification with version numbers is one you can pin against, cite in a compatibility table, and reason about when an agent platform claims to support Agent Skills. Without numbers, that claim means whatever the vendor read most recently.

A formal proposal process is what turns a GitHub organisation from a mailing address into a decision mechanism. The hard part is never the existence of the process; it is the rules. Who may propose, who decides, what happens when a proposal is rejected, and whether a rejection is published are all unanswered in the announcement, which is why proposed is doing real work in that sentence.

Security guidance is the item with teeth, and it is the one to watch first. The format allows a skill to bundle instructions together with optional resources, and those resources can include executable scripts. Installing a skill is therefore not a purely reading act; it is closer to installing a small piece of software. Guidance aimed at both people who write skills and people who install them covers the two ends of that transaction, which is the right place to start and the hardest place to get agreement.

Read together, the three items describe a specification growing up: numbers so it can be depended on, a process so it can change without a fight, and guidance so it can be installed without an incident. Read together, they also describe a project with a substantial amount of work still ahead of it.

Seventy platforms is the number that reframes the announcement

According to Anthropic's news index summary, more than 70 agent platforms have integrated Agent Skills since the standard was published in December 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code. That is a vendor-reported figure and should be read as one: it counts platforms that integrated the format, not people using it, and it says nothing about how deep any individual integration goes.

Take it as a floor, and the governance move becomes legible. A format integrated in seventy places has one characteristic failure mode, and it is not obscurity. It is divergence: every ambiguity in the specification becomes a place where two implementations can quietly disagree, and the disagreement surfaces only when somebody moves a skill folder between agents and gets two different results.

A foundation with versioned releases is a response to that, and it is a response to a problem a single vendor is badly placed to solve alone. The company that wrote a specification cannot credibly arbitrate between its own implementation and everyone else's, however fair it tries to be, because the appearance of the conflict is enough to erode the standard. A body whose maintainers include engineers from several adopters can. That is the strongest argument in the announcement, and Anthropic makes it implicitly.

It also explains a detail that otherwise looks odd. OpenAI appears twice in the story, once as a co-founder of the foundation and once as an employer of maintainers. For a standard whose job is to let the same folder run under agents built by different companies, that overlap is the design.

What to watch now that the post is over

The specification lives at agentskills.io and the feedback channel is the project's Discord. Those two addresses are the record from here on, and they are a better place to look than the announcement, which states intent rather than documenting process.

Three things would confirm that the transfer changed something real rather than relabelled something that already worked. The first is a version number on the specification, because a version number is the artefact that a commitment to versioned releases produces. The second is a proposal that gets turned down in public, since a process that has never said no is not yet a process — it is a channel. The third is security guidance actually shipping in some form, because it is the item that touches people who install skills they did not write.

Until one of those lands, the accurate description of Agent Skills is this: the same format, a different owner of its name, maintained by named individuals under a process that has been proposed but not yet described. That is a meaningful change and a small one at once, which is why commentary on it keeps splitting into the two readings above.

There is a reason a change like this matters beyond governance. OrcaRouter routes model calls — more than 200 models behind one OpenAI-compatible key, provider list price passed through at zero markup, automatic failover between providers — and we host none of Anthropic's apps or its skill surfaces. A skill folder that a team writes once and expects to run unchanged under agents from several vendors is the same kind of argument as a model call that should not care which provider answers it. Standards earn their keep on exactly that indifference, and so does a router.

The donation did not make Agent Skills better. It made it harder to take back. For anyone with a skill folder sitting in a repository, that is the whole of the benefit, and it is not a small one: the format you built on now has an owner whose interest is in the format surviving, and a maintainer list that spans more than one company. The next version number will tell you whether the rest follows.