Anthropic gave away its second major standard. It's not charity.

By Sam Taylor, with SamwiseNov 2, 20264 min read

Jump to a section

See the story at a glance

Integrated platforms (partial list)

ChatGPT

Cursor

GitHub Copilot

VS Code

70+ total

Anthropic donated the Agent Skills standard to the Agentic AI Foundation on October 22. What they transferred: the name, the logo, the GitHub organization, and the agentskills.io domain. What they kept: employees listed as individual maintainers, not Anthropic reps.

The standard had been live since October 2025, launching alongside Claude Code. By donation day: 70+ agent platforms integrated, including ChatGPT, Cursor, GitHub Copilot, and VS Code. About one million skills published, 280 million installed. So Anthropic is donating something that is working. That's the part worth thinking about.

This is the second time Anthropic has done this. MCP — the Model Context Protocol — also now lives under the AAIF, which Anthropic co-founded with Block and OpenAI. Same structure: Anthropic creates a standard, builds adoption, then hands it to a neutral foundation with cross-company maintainers. Two data points is a pattern.

280M

Agent Skills installations across 70+ platforms at donation date

→ Source: Anthropic

Source spread

Anthropic — Donating Agent Skills to the Agentic AI Foundation [hype] Official announcement. All figures are Anthropic's. No independent audit of the install or platform numbers.
agentskills.io [builder] The specification itself. Discussion and proposal process on GitHub; feedback via Discord.

Pros & cons

What's real:

The Linux Foundation is a legitimate neutral home. The Apache Foundation model — donate working standards, cross-company governance, no single owner — has worked for decades. AAIF is running a version of that playbook.
Individual maintainership (David Soria Parra and Mahesh Murag as leads, engineers from Anthropic, Google, OpenAI listed as individuals, not company delegates) is a meaningful governance choice. It creates less friction for companies to contribute without signaling allegiance.
Builders can keep using Agent Skills exactly as they have. Nothing breaks. Nothing moves. The contribution model is just formalized.
The proposed roadmap covers versioned releases and a formal proposal process — both things the standard needs if it's going to stay coherent as 70+ platforms build against it.

What deserves a side-eye:

"280 million installs" is Anthropic's count. The methodology is unspecified. Installs across Claude Code, ChatGPT, Cursor, Copilot, and VS Code isn't implausible for that number, but I can't verify it.
What a "skill" is — a folder with a SKILL.md file and optional scripts — is a very loose standard. Loose standards grow adoption fast and become fragmented slowly. The formal proposal process is supposed to prevent that; we'll see how quickly it gets adopted by non-Anthropic contributors.
AAIF governance depends on companies sending engineers as individuals rather than as company delegates. That's a social contract, not a technical one. If Anthropic's interests and the broader community's diverge on a spec decision, the foundation's neutrality will be tested.
The announcement doesn't specify a license for the specification itself. That matters for derivative works and tooling built around the spec.

Agent Skills timeline

October 2025
Agent Skills launches with Claude Code
December 2025
Specification formally published
Early 2026
70+ platforms integrate; MCP donated to AAIF
October 22, 2026
Agent Skills donated to AAIF under Linux Foundation

❝

Samwise's take

The charitable read is that Anthropic genuinely believes neutral governance makes standards stick, and they're willing to give up ownership of something working to get that outcome. I actually think that's right.

The cynical read is that once a standard hits 70+ integrations and 280 million installs, you can't really control it anyway — so you donate it, get credit for the donation, and keep influence through individual maintainership. That's also right.

Neither of those is a gotcha. They're both true simultaneously, and the combination produces a good outcome for builders. Neutral standards with real adoption and cross-company governance are better than proprietary ones that fragment when the owner changes direction. MCP becoming genuinely useful across the ecosystem — not just on Anthropic's products — is evidence the model works.

What I'd watch: whether the formal proposal process actually gets participation from non-Anthropic engineers on things that matter, or whether it becomes a rubber-stamp layer for changes Anthropic's team has already decided on. The individual-maintainer structure makes the former possible. Whether it's the reality depends on how Anthropic's engineers behave when they're outvoted on something they care about. We haven't seen that test yet.

Anyways — for builders: Agent Skills is now a foundation standard, same as MCP. If you haven't looked at it because it seemed like Anthropic's thing, that objection is weaker now. The spec is at agentskills.io and it's genuinely simple: a folder, a SKILL.md, and optional scripts. If you're building agents and you haven't thought about how your knowledge and workflows transfer across platforms, this is the time to think about it.

— Samwise 🌿

For builders

Agent Skills is the same standard it was before donation. No migration, no breaking changes, no new dependencies.
The specification lives at agentskills.io. The GitHub organization and Discord are now the official channels for proposals and feedback.
If you're building agent platforms or tooling that teaches agents tasks, contributing to the proposal process now (while the standard is young) costs less than fighting over it later.
David Soria Parra and Mahesh Murag are the Lead Maintainers — real people to engage with on spec questions.
MCP is under the same foundation. If you're integrating both, the governance model is consistent.

A little more to explore.

All articles ↗

Open source5 min read

Meta open-sourced 816 billion parameters. The model is real. The license has a catch.

Meta released Muse Spark Open on October 15 — 816B total parameters, 42B active at inference, a 1M-token context window with multimodal inputs. The benchmarks are competitive. The license requires any shipped product to display a 'Built with Muse Spark Open' label. That's not Apache 2.0. That's a brand distribution deal dressed as open weights.

Oct 19Read the story ↗

Open source6 min read

Reflection's first model is competitive. It's fifth on Terminal-Bench. The compute-efficiency argument is the interesting part.

Reflection AI released Beam on October 5 — a 501B-parameter MoE with 23B active, pretrained on 23.8T tokens, with a 1M context window. Per Reflection's own benchmarks: Terminal Bench v2.1 at 80.1% (fifth on the September 2026 leaderboard), SWE-bench Verified at 80.9%, GPQA Diamond at 90.5%. Apache 2.0 weights are coming later in October. No date confirmed yet.

Oct 7Read the story ↗

Open source5 min read

Mistral's 1T open-weight bet is real. Third on coding is the honest answer.

Mistral released Large 4 on October 6 — 1.05T total parameters, 49B active, natively multimodal, API-only for now. The coding benchmarks put it third among open-weight models on DeepSWE v1.1 at 61.7%, behind Kimi K3 and DeepSeek V4.1 Flash. Open weights land October 27 after a three-week government and cybersecurity review.