Claude Agent Skills Join the Agentic AI Foundation: Standardize or Wait?

On October 22, 2026, Anthropic donated the Agent Skills name, logo, GitHub organization, and agentskills.io domain to the Agentic AI Foundation, the Linux Foundation body that already hosts the Model Context Protocol. If your team packages agent capabilities, the practical answer is: adopt skills where Claude is the runtime and the payload is instructions plus scripts, keep tool transport in MCP servers, gate every shared skill on security scans and live evaluation, and hold off on org-wide standardization until the foundation’s versioned-release and proposal processes actually exist. The donation transfers assets, not a compatibility guarantee.

That distinction matters more than the headline. The announcement is real and the governance intent is genuine, but nearly everything a platform engineer needs to bet a shared skill library on, the charter, the license terms, the shipped versioning process, and verified non-Anthropic runtime commitments, is either proposed or unverified as of this writing.

What the donation actually transfers

Anthropic’s announcement is unusually precise about scope, and the precision cuts both ways:

“We are donating the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain. You can continue to use the standard or contribute to it without joining the AAIF.”

Four assets move: a trademark pair, a code organization, and a domain. What does not move is equally clear. There is no transfer of a conformance test suite, no registry, and no runtime commitment from any non-Anthropic vendor named in the announcement. The foundation itself is a Linux Foundation body Anthropic co-founded with Block and OpenAI, the same home that received MCP, so the institutional path is proven. But proven institutional path and proven standard are different claims.

On maintenance, the announcement says engineers from Anthropic, Google, OpenAI, and other companies will maintain the standard “as individuals,” with David Soria Parra and Mahesh Murag as Lead Maintainers. Individual maintainership is common in healthy standards work, but it is weaker than corporate stewardship: employer priorities can pull an individual maintainer away with no organizational obligation to replace them. And the governance roadmap is explicitly a proposal:

“We’ve proposed starting with a roadmap covering versioned releases, a formal proposal process, and security guidance for people who write and install skills.”

Proposed is not adopted. Until versioned releases ship, there is no stable target to pin a skill library against, and no formal process by which your team could contest a spec change that breaks you. The announcement does not publish the AAIF charter or its trademark-use terms; confirm both before treating any of this as settled. The specification license is documented separately: the GitSkills dataset paper’s reference list cites the Agent Skills specification as Apache-2.0.

How skills and MCP split the packaging layer

The two specifications now sitting in the same foundation solve different problems, and conflating them is the most common design error here.

MCP is transport: a protocol for an agent to discover and invoke tools hosted by servers, with the tool’s implementation living behind the protocol boundary. Agent Skills are packaging: a directory format, introduced by Anthropic in October 2025 with its structure, front-matter constraints, and staged loading model published as an open specification, per the GitSkills dataset paper. A skill is a folder: a SKILL.md file with instructions, plus optional scripts and resources that the agent loads progressively rather than paying full context cost up front.

The practical split:

A skill that wraps an MCP tool is a reasonable pattern; a skill that re-implements what an MCP server already exposes is duplication with worse provenance. If your organization already runs MCP servers for tool access, skills do not replace them. They sit above them, packaging the procedural knowledge of when and how to use those tools.

The portability record: 70 integrations is not interchangeability

Anthropic reports that more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have integrated Agent Skills since the standard was published in December 2025. Treat that count as vendor-reported: it comes from the donating party’s own announcement, and “integrated” is doing quiet work in that sentence. Reading a SKILL.md file and rendering its instructions is a much lower bar than executing a skill’s bundled scripts with the same semantics the author’s runtime provided.

The strongest independent counterweight is the skills survey arXiv:2602.12430, which concludes:

“Achieving genuine portability requires either (a) a universal skill runtime or (b) skill compilation targeting different agent platforms.”

The survey’s reasoning is that skills authored for Claude may implicitly depend on Claude-specific capabilities such as the code execution environment, tool signatures, and model behaviors. A skill whose instructions assume a particular tool set, file layout, or sub-agent mechanism will parse fine on another platform and behave differently or fail quietly. That is the failure mode that matters for a shared library decision: not a loud incompatibility, but a skill that loads everywhere and works in one place.

This is an analytic conclusion from a survey, not a runtime-by-runtime compatibility matrix, so it cannot tell you which specific skills port and which do not. What it does is move the burden of proof. Before standardizing cross-runtime, you need your own portability trials on the skills you actually intend to share, not a platform count from a press release.

Supply-chain math: copies of copies

The GitSkills dataset gives the ecosystem its first measured baseline: 3,797,117 SKILL.md files across 282,200 public GitHub repositories owned by 195,841 accounts, collected in July 2026. Adoption is genuinely broad. The reuse mechanism, however, is primitive:

“Skills have no central registry; reuse happens by copying folders, and 50.5% of the collected files are byte-for-byte copies of another file in the dataset.”

Half the ecosystem is duplicates with no version pointer, no update channel, and no way to answer “where did this skill come from and has it been fixed?” Anyone who has chased a copied-pasted config file through an organization will recognize the shape of the problem. The donation does not address it; no registry appears among the transferred assets or the proposed roadmap items.

Security sharpens the concern. The skills survey cites a scan of 42,447 community skills reporting a 26.1% vulnerability rate, and argues the current implicit trust model is untenable. Two caveats belong with that number: it is cited within the survey from separate work (Liu and others, 2026a), not measured by the survey itself, and “vulnerability” will depend on the scanner’s definitions. Even discounted, a one-in-four-ish figure on artifacts that are typically installed by folder copy, directly into an agent’s execution context, is enough to make “we only install skills we scanned” a minimum bar rather than a best practice.

What a growing skill library costs to run

Even setting portability aside, more skills is not monotonically better, and the measured curves are steeper than intuition suggests.

The ACES evaluation framework, which tested 145 real skills from enterprise repositories and public catalogs, found that scan-only gates surface useful authoring issues but measure facets complementary to live trials; you need both. Its stress study, per the ACES paper, is the number to put in front of anyone proposing a large shared catalog:

“Mean overall lift stays at 0.133–0.149 for 1–20 visible skills while mean wall time rises from 258 seconds at one to 451 at 20. At 50, the with-skill pass rate is 0.55 and mean wall time is 1,290 seconds”

Read that as: benefit plateaus early, latency grows roughly linearly, and past some catalog size the agent gets worse at the task than it would be with fewer skills. These are author-reported preprint results from a specific 25-variant study; treat the shape as informative and the exact figures as benchmark-bound.

Retrieval architecture is the main lever. On Graph-of-Skills, dependency-aware structural retrieval beat vector retrieval by 10.97 points on SkillsBench and 2.87 points on ALFWorld, averaged over three model families. Under Claude Sonnet 4.5 specifically, the authors report 97.9% average success on ALFWorld versus 93.6% for vector retrieval, while cutting average total tokens from 1,524,401 to 27,215. Again, author-reported on specific benchmarks, but the mechanism is plausible and general: when the library is large, how you pick skills dominates what the skills contain.

One more operational finding worth baking into process: ClawTrace reports that stripping cost attribution from its tracing cards during skill distillation more than doubled the median cost uplift on successful tasks, from +22% to +49%. In plain terms, if your pipeline for creating and refining skills does not record what each skill costs to run, you risk distilling skills that quietly inflate spend. Cost accounting belongs inside the skill lifecycle, not in a quarterly review afterward.

The diligence checklist

Before your organization treats Agent Skills as a standardization target, these are the checks that separate a real foundation transition from a branding one:

What would flip the verdict

My recommendation is a staged posture, and it is worth stating the conditions that change it, because they are observable rather than vibes-based.

Standardize org-wide on Agent Skills when three things exist: AAIF has shipped versioned releases and an adopted proposal process, the charter and license terms are published and confirmed acceptable to your counsel, and at least one runtime you depend on besides Claude has committed to the spec in a form you can test. At that point the compatibility burden genuinely moves from one vendor’s roadmap to a governance process, and a shared cross-runtime skill library becomes a rational investment rather than a bet.

Until then, the format is useful today under narrower conditions. If Claude is your runtime, skills are the natural packaging for instructions-plus-scripts work, the staged loading model is well suited to keeping context costs down, and the December 2025 spec is public and usable without joining anything. Keep tool transport in MCP. Curate an internal catalog rather than pulling from community folders, scan and live-test everything in it, track per-skill cost, and cap the visible skill count well below the point where the ACES curve turned. And if a lost week of migration work would hurt, that is a reason to wait, not a reason to move faster: the assets being donated will still be there when the governance catches up to them.

Frequently Asked Questions

What specific assets did Anthropic donate to the Agentic AI Foundation?

We are donating the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain. You can continue to use the standard or contribute to it without joining the AAIF.

What is the recommended approach for organizations using Agent Skills today?

If Claude is your runtime, skills are the natural packaging for instructions-plus-scripts work, the staged loading model is well suited to keeping context costs down, and the December 2025 spec is public and usable without joining anything. Keep tool transport in MCP. Curate an internal catalog rather than pulling from community folders, scan and live-test everything in it, track per-skill cost, and cap the visible skill count well below the point where the ACES curve turned.

What conditions must be met to standardize Agent Skills organization-wide?

Standardize org-wide on Agent Skills when three things exist: AAIF has shipped versioned releases and an adopted proposal process, the charter and license terms are published and confirmed acceptable to your counsel, and at least one runtime you depend on besides Claude has committed to the spec in a form you can test.