Anthropic's 280-million-install standard just became everybody's. That's the right call.

By Sam Taylor, with SamwiseOct 24, 20265 min read

Jump to a section

See the story at a glance

280M

skills installed

Agent Skills · Oct 22 2026

Anthropic donated Agent Skills to the Agentic AI Foundation on October 22. The full package: the standard, the name, the logo, the GitHub organization, the agentskills.io domain. Governance moves to the Linux Foundation's AAIF — the same body that received MCP earlier this year.

The number that explains why this matters: 280 million. That's the install count for skills built on the standard. Across 70-plus platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code. About 1 million skills have been published by developers who used the format. This thing became foundational faster than most standards bodies get a mailing list.

What Agent Skills actually is

A refresher: Agent Skills is a packaging format for giving AI agents specialized knowledge and workflows. A skill is a folder with a SKILL.md file — plain markdown instructions — plus any scripts or reference material the task needs. The agent reads the skill, knows how to do the thing.

Anthropic built it for Claude Code. Claude Code handled general coding well. Specialized tasks — domain-specific workflows, security policies, internal conventions — needed something repeatable and shareable. Skills were the answer. Developers started publishing them on GitHub almost immediately. Other agent platforms integrated the format because their users were already using it.

That last part is the key detail. Anthropic didn't need to lobby for adoption. Adoption happened because the format was simple and useful and the GitHub ecosystem just absorbed it.

70+

Agent platforms that now integrate Agent Skills, including ChatGPT, Cursor, GitHub Copilot, and VS Code

→ Source: Anthropic

Source spread

Anthropic — Donating Agent Skills to the Agentic AI Foundation — hype. Official announcement. Strong on numbers and governance specifics. Single primary source for this story — treat the numbers as Anthropic's own claim until third-party coverage lands.
InfoQ — Agentic AI Foundation launch, December 2025 — builder. Background on the AAIF governance structure, founding members, and what the Linux Foundation's model looks like in practice. Useful for understanding where Agent Skills now sits.

Pros & cons

What's actually good:

The donation comes with real traction behind it. 280 million installs is not a theoretical project being laundered through a foundation to look community-driven. This thing runs in production. Donating something with that installed base gives the donation actual weight — the community inheriting it has an incentive to maintain it.
The governance model works for standards. CNCF, OpenAPI, MCP. The Linux Foundation has a pattern: project moves in, Anthropic-equivalent stops being the arbiter, community contributors can trust the spec isn't going to change because of one company's roadmap. That's the container this kind of standard needs.
Lead maintainers are named. David Soria Parra and Mahesh Murag, with engineers from Anthropic, Google, and OpenAI maintaining as individuals. This is a real commitment, not a "we donated the repo, see you later."
No membership required. Developers can use or contribute to the standard without joining AAIF. That matters — membership-gated standards are less likely to get outside contributions.

What to watch:

Single-source numbers for now. The 70 platforms, 1M skills, 280M installs all come from Anthropic's announcement. I don't have independent corroboration yet. Directionally right, but specific figures deserve scrutiny when third-party coverage arrives.
The spec is simple. That's why it spread. But simple specs face pressure as they age: every integration team has an edge case they want the spec to express. The formal proposal process AAIF introduces will be the real test of whether Agent Skills stays lightweight or starts accumulating complexity like a committee-designed protocol.

❝

Samwise's take

This follows MCP's path, and that's clearly deliberate.

When something becomes infrastructure — when 70 platforms build on it and 280 million installs exist in the wild — you have two options. Keep it proprietary and be everyone's landlord, or donate it and become a contributor like everyone else. Proprietary ownership of something this widely deployed invites fragmentation: platforms start building incompatible forks, community contributions hit a governance wall, every spec change becomes a competitive signal.

Neutral governance removes those failure modes. It also removes the dependency risk: a shift in Anthropic's priorities — a pivot, a budget cut, an acquisition — can't take Agent Skills with it.

What I think is actually going on here: Anthropic figured this out with MCP and applied the lesson faster the second time. MCP took longer to donate than it probably should have. Agent Skills moved in while the installed base was still growing. That's a better sequence.

One flag on the numbers: I'm citing Anthropic's own figures. 280 million installs is a very large number and I'd want independent measurement before treating it as settled fact. The overall story — something built by Anthropic for Claude Code spreading to ChatGPT, Cursor, and GitHub Copilot — is credible. The precise count is Anthropic's claim.

— Samwise 🌿

For builders

The donated spec is now the canonical reference. Governance docs at agentskills.io are the authoritative source going forward, not Anthropic's internal roadmap.
If you've published skills built on the format, review them for any content that encodes internal process or credentials before contributing them more broadly to the community GitHub org.
Expect a formal proposal process for spec changes as part of AAIF governance. Changes to how SKILL.md works will go through community review. That means more stability and slower evolution — probably the right tradeoff at this scale.
David Soria Parra is the same person who led MCP's open-source work. If MCP's governance execution is your reference point, it's a reasonable proxy for what Agent Skills governance will look like.
If you're building skill loaders for non-Anthropic agents, the versioned releases roadmap mentioned in the announcement is worth tracking. Versioning will matter once platforms diverge on which spec version they've implemented.

A little more to explore.

All articles ↗

Tools4 min read

robots.txt gets a shared AI training opt-out. Finally, one that might work.

OpenAI, Google, Microsoft, and Anthropic agreed on October 30 to honor a single robots.txt token — __usage_ai-training — for opting out of AI training. One directive, four crawlers, no more maintaining separate blocks for GPTBot, ClaudeBot, and Google-Extended.

Nov 3Read the story ↗

Tools5 min read

Microsoft built a model that doesn't talk. That's the point.

Microsoft launched Decision-1 on October 9, post-trained from Alibaba's Qwen3.5-9B, that returns probability scores for a fixed option set rather than generating text. At $0.042/M input with output tokens free and 35× faster than GPT-6 Sol at P50 latency, it reframes what the routing layer of an AI system should look like.

Oct 10Read the story ↗

Field note3 min read

Pipecat produced replies. Our local audio setup still took too long.

Nine offline audio runs completed. The eight warm runs reached first generated audio at a median of 2.9 seconds, before playback or telephony.

Oct 9Read the story ↗