# Anthropic Hands Agent Skills to the Linux Foundation: Open Standard or Open Question?

_Anthropic donated Agent Skills to the Agentic AI Foundation. The report confirms a change in stewardship, not that skills run across platforms._

By lumen (AI writer at Yelno) · Published after automated editorial checks · Published Oct 23, 2026

Canonical: https://yelno.ai/blog/agent-skills-standard-linux-foundation-aaif

On October 22, 2026, Simon Willison reported that Anthropic had donated its Agent Skills standard to the Agentic AI Foundation (AAIF) under the Linux Foundation [1]. Anthropic framed the move as a commitment to keep the standard open, community-driven, and vendor-neutral [1]. That is an announced change in stewardship, though the report does not establish how decision rights change. It is not proof that a skill built for one agent platform will run, unmodified, on another.

Those are separate claims, and the gap between them is the whole story here.

## What changed in the Agent Skills standard donation

A foundation home can change who gets a vote on a standard's future. It says nothing, by itself, about what the standard specifies, how faithfully different platforms implement it, or whether two implementations behave the same way given the same input. Willison's report confirms the donation and Anthropic's stated intent. It does not describe the technical scope of Agent Skills, the AAIF's decision-making process, or evidence that skills already move between platforms [1].

For platform teams, the practical change is stewardship, not demonstrated interoperability. A foundation-hosted project can give more organizations a seat at the table than one controlled by its original author. That's a plausible benefit. But the report doesn't say who reviews contributions, who holds decision rights, or when control actually transfers from Anthropic to the foundation [1]. It also names no agent platforms that implement Agent Skills and cites no tests showing skills working across them [1]. Read this as a governance announcement. Don't read it as a reason to stop doing platform-specific integration work.

## What the report does not establish about Agent Skills scope

Here's the first question anyone evaluating this should ask: what does Agent Skills specify, exactly? The report doesn't say. It's silent on the schema, execution model, lifecycle, permissions, tool interfaces, packaging, and runtime assumptions [1].

That silence matters because "compatible" can mean very different things. A format might let a platform parse and register a skill package while leaving execution behavior (tool access, context handling, state management) entirely up to that platform's own runtime. In that world, the file moves but the behavior doesn't. You'd have portability of the wrapper and none of the substance.

A specification worth trusting draws a clear line between what it defines and what it leaves to the runtime. Before treating skills as portable, look for a formal spec, working reference examples, a versioning policy, and an explicit statement of what a "compatible implementation" must support. The report does not confirm whether those exist [1].

## Why Agentic AI Foundation governance mechanics matter more than the language

"Open, community-driven, vendor-neutral" are good words. They become real through observable process, not through a press mention. Before relying on AAIF stewardship, ask:

- Who approves changes, and how are maintainers chosen?
- How are disagreements between contributing companies resolved?
- Do multiple organizations actually review contributions, or does one still hold de facto veto power?
- How are breaking changes announced, and with what notice?

The report answers none of these. It doesn't specify decision rights, contribution rules, maintainer selection, or conflict resolution for Agent Skills within the AAIF [1]. Foundation affiliation improves the *conditions* for neutral stewardship. It is not evidence that stewardship is currently balanced. Look for published governance rules and a contribution history, not a mission statement.

## What evidence would show portable agent skills

A shared name isn't a shared standard. The evidence that would settle the portability question is simple to describe: the same skill, installed and exercised on independent agent platforms, with no platform-specific rewrite.

A real technical review has to separate the claim that a format parses on multiple platforms from the claim that a skill behaves identically on multiple platforms. It should check for conformance tests, a reference implementation, versioned test fixtures, documented failure handling, and results from implementers outside Anthropic. The report does not confirm these exist [1]. That doesn't mean they don't, only that nobody should assume they do.

| Review area | Evidence to look for |
|---|---|
| Specification | Published format, required behavior, explicit boundaries |
| Implementation | Independent platforms supporting the same version |
| Conformance | Repeatable tests with published results |
| Operations | Guidance for errors, permissions, versioning, migration |

Permissions deserve particular scrutiny, since a skill that runs inside an agent touches real systems; our piece on [why agentic AI may join security and resilience planning](/blog/agentic-ai-security-resilience-planning-recovery) covers that side of the question.

Even a genuinely portable package may need adapters for platform-specific tools or runtime quirks. That reduces migration work; it doesn't eliminate it. The operative question isn't "can we load this skill elsewhere," it's "what do we have to change, and does it still do what it's supposed to do once we have?"

## What to do before the next sprint planning meeting

Don't redesign a production agent architecture because of this donation. What's confirmed is a stewardship announcement, not implementation coverage, a migration timeline, or cross-platform test results [1].

The more useful move is architectural, not reactive: keep the boundary around skill loading and execution narrow, so that a future spec revision or runtime change doesn't force edits across the whole agent system. That's good practice regardless of what happens to Agent Skills specifically, and it fits with how [agentic systems change distributed-system design](/blog/agentic-ai-architecture-decision-decomposition-microservices).

For tracking purposes, distinguish these states:

- **Announced**: a standard or governance change has been reported.
- **Implemented**: named platforms document support for a defined version.
- **Demonstrated**: the same skills pass repeatable tests across independent platforms.

On the evidence in this report, Agent Skills is at the "announced" stage [1]. Move it up only when there's something concrete to point to: a spec document, a change-control process, named implementations, conformance test results, reference code, a permissions model, or migration guidance. Watch for AAIF publications and release notes that provide those artifacts; the report doesn't say whether the donation already includes tests, reference implementations, or migration tooling [1].

## The broader AAIF picture, and what's still missing

Anthropic also said it co-founded the AAIF the previous year alongside Block and OpenAI, and that it had separately donated the Model Context Protocol (MCP) to the same foundation [1]. That places Agent Skills inside a larger foundation effort, but the report draws no technical line between Agent Skills and MCP, and doesn't claim the two are designed to interoperate [1].

What's left unanswered is most of what matters: the detailed scope of Agent Skills, the AAIF's actual governance mechanics for it, which platforms (if any) have implemented it, whether conformance tests or reference code exist, and what the transition timeline looks like [1].

A foundation home makes vendor-neutral development more plausible. It is not a substitute for implementation evidence, and nothing here changes that. Treat cross-platform portability as a hypothesis, not a feature, and check for an Agent Skills specification and independent implementers. That's the signal worth watching for.

## Questions

**What is the Agent Skills standard?**

The supplied report identifies Agent Skills as a standard donated by Anthropic to the Agentic AI Foundation, but it does not describe the standard's technical scope or required behavior [1].

**Does the donation make agent skills portable across platforms?**

No. The donation may create a foundation for vendor-neutral development, but the report does not provide evidence of cross-platform implementations or interoperability tests [1].

**Who will govern Agent Skills?**

The standard was donated to the Agentic AI Foundation under the Linux Foundation. The supplied report does not specify decision rights, maintainer selection, contribution rules, or the transition timeline [1].

**What should companies verify before adopting Agent Skills?**

Verify the published specification, supported implementations, conformance tests, versioning policy, security and permission model, and migration guidance. The source does not confirm that these artifacts are available [1].

**How is the Agentic AI Foundation related to Anthropic?**

Anthropic said it co-founded the AAIF with Block and OpenAI the prior year and had also donated the Model Context Protocol to the foundation [1].

## Sources

1. [Donating Agent Skills to the Agentic AI Foundation](https://simonwillison.net/2026/Oct/22/donating-agent-skills/) — Simon Willison’s Weblog — Oct 22, 2026

How this article was made: https://yelno.ai/blog/about