Anthropic Just Gave Away Agent Skills — and Three Rivals Gave Away Pieces of the Agent Stack With It
The most useful thing Anthropic built this year was a folder
Not a model. Not a benchmark score. A folder — containing one markdown file called SKILL.md, plus any scripts and reference material an agent needs to do a specific job properly. On October 22, at AGNTCon + MCPCon North America in San Jose, Anthropic donated that format, Agent Skills, to the Agentic AI Foundation (AAIF) under the Linux Foundation — transferring the name, the logo, the GitHub organization and the agentskills.io domain with it.
It did not come alone. The same announcement added three more projects from three other companies: Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. Together they take the AAIF from its three founding projects in December 2025 to 10 hosted projects in its first year. Four competitors, on one stage, handing pieces of their agent technology to a neutral home so that, in the foundation’s words, no single company controls the outcome.
That is either a genuine turning point for how agents get built — or the most coordinated positioning the industry has produced this year. Probably both. Either way, builders should understand what changed, because skills are already inside the tools most teams use.
What a skill actually is, and why it spread
Anthropic developed skills while building Claude Code. The model handled general knowledge work well, the company says, but lacked the procedural knowledge real business tasks need — the house way of filling a spreadsheet, reviewing a contract, processing a PDF, or running a particular workflow in the right order with the right checks.
A skill fixes that without retraining anything. It is a directory: a required SKILL.md file carrying a short name and description plus step-by-step instructions, and optional folders for scripts the agent can run, references it can read on demand, and assets such as templates or data files. The agent reads the description, decides the skill is relevant, and loads the full instructions only at that moment — so a library of skills costs almost nothing until one is needed.
The simplicity is the strategy. There is no runtime to install, no package manager, no registry account. If an agent can read files, it can in principle use a skill. Within weeks of the October 2025 launch in Claude Code, developers had published thousands of skills on GitHub and written loaders for Codex, Cursor and OpenCode — Anthropic’s competitors’ tools, loading Anthropic’s format voluntarily. In December 2025 Anthropic published Agent Skills as an open standard under its own organization and domain, so any agent could use it.
The adoption numbers are Anthropic’s own, not independently audited — but even discounted, they describe a standard that escaped its creator: more than 70 platforms have integrated Agent Skills, including ChatGPT, Cursor, GitHub Copilot and VS Code; 1 million skills published; nearly 280 million installed.
Why donate something that is winning?
That is the question worth asking, because companies rarely give away a format that is beating the alternatives. The pattern, though, is familiar from earlier platform wars: a format’s value rises with the number of products that support it, and rival products support it faster when they no longer worry that its owner will change the rules, charge for it, or tilt it toward its own models. Donating to a foundation is how an owner credibly promises that will not happen.
Anthropic has run this play before, recently. It co-founded the AAIF with Block and OpenAI last year and donated the Model Context Protocol (MCP) — the standard for connecting agents to external tools and data — to it less than a year ago. The division of labour between the two donations is clean: MCP is how an agent reaches tools and data; Skills are how it knows what to do with them. Anthropic’s David Soria Parra, a co-creator of MCP and now a lead maintainer here, put the pitch plainly: “Without the right context, even the most intelligent agents fall short.”
Governance specifics matter more than the ceremony. Engineers from Anthropic, Google, OpenAI and other companies will maintain Agent Skills as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. Anyone can keep using or contributing to the standard without joining the AAIF. Anthropic has proposed a starting roadmap covering versioned releases, a formal proposal process, and security guidance for people who write and install skills. Versioning and a real proposal process are unglamorous — and they are exactly what an enterprise needs before it builds internal workflows on a format.
The other three donations complete the stack
Read the four projects together and they map, neatly, onto the whole agent stack:
- Agent Skills (Anthropic) — how agents gain new capabilities: portable know-how they load on demand.
- Open Responses (OpenAI) — how applications talk to models: an open specification based on the OpenAI Responses API, with a shared schema, consistent streaming events and extensible tooling, so a request written once can run across model providers with minimal translation.
- OpenEnv (Hugging Face) — how agents are trained and evaluated: a framework and hub for reinforcement-learning environments, delivered as a containerized service exposing an agent interface, MCP and the Gym training protocol, backed by a committee spanning Hugging Face, Microsoft, Reflection, Meta-PyTorch, Fleet, Mercor, RadixArk, Modal and Prime Intellect.
- OpenSandbox (Alibaba) — where agents run safely: a general-purpose sandbox platform with multi-language SDKs, unified APIs and Docker and Kubernetes runtimes for coding agents, GUI agents, evaluation and code execution, plus a protocol for custom runtimes.
They join A2A, Agent Router, agentgateway, AGENTS.md, goose and MCP. AAIF CTO Manik Surtani framed the milestone directly: “A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10.” Alibaba’s Xiaobin Xu said the quiet part for his company too: OpenSandbox was created at Alibaba, but “we don’t want its future to be defined by a single company.”
The risk a neutral home makes bigger, not smaller
A shared standard spreads good skills fast. It spreads bad ones at the same speed, and the roadmap’s promise of security guidance is an admission that this part is not solved. A skill is not just text an agent reads — it can carry scripts an agent runs, and instructions an agent follows. Installing a skill from a stranger is closer to installing software than to reading a document.
That concern is not theoretical. Secondary reporting noted Snyk confirmed 76 malicious skills in samples from the ClawHub directory — treat that figure as reported, not independently verified here — and this week Claude Code 2.1.309 closed a real sandbox escape in which sandboxed commands could run shell startup files outside the sandbox. Agents that load portable instructions and then execute code need both halves hardened. It is telling that AAIF now hosts a skills standard and a sandbox project announced the same day.
The practical rule for teams is unchanged, and stricter under a neutral standard because sharing gets easier: read a skill before you install it, audit any scripts it tells the agent to run, pin versions from sources you trust, and treat an unknown skill the way you would treat an unknown dependency — because that is what it is.
What to do with this, this week
- Nothing breaks today. No SKILL.md file has to change because of the donation. If your team already writes skills for Claude Code, ChatGPT, Cursor or Copilot, the format continues — now with versioned releases promised instead of one company’s changelog.
- Write skills once, deliberately. The portability case is now real across 70-plus platforms. The internal procedures your agents repeat — document handling, reviews, reporting, data checks — are worth capturing as maintained skills rather than re-prompting from scratch each time.
- Add a skills review step. Before any third-party skill enters your environment, review its instructions and bundled scripts, check who maintains it, and sandbox its first runs. Put that step in writing; the standard’s own roadmap is heading there.
- Watch Open Responses if lock-in worries you. A provider-neutral interface spec could make switching models a configuration change instead of a rewrite. Judge it by provider implementations, not the announcement.
The bottom line
Models still get the headlines, but durable power in AI lives one layer down, in the formats everyone else builds on. On October 22, four companies decided that layer for agents — skills, model interfaces, training environments and sandboxes — should live somewhere none of them own. Anthropic gave away the fastest-spreading piece because a standard nobody fears gets adopted by everybody. Whether AAIF governance stays genuinely neutral is a claim to verify over the next year. The direction is set: the agent stack is becoming shared infrastructure, and competition is moving back to the models and products on top of it.