On October 22, 2026, Anthropic donated Agent Skills, its open format for packaging reusable agent know-how, to the Agentic AI Foundation under the Linux Foundation. It is the second Anthropic-born standard to get a neutral home after the Model Context Protocol, and it completes a protocol stack, now ten projects strong, that governs how AI agents talk to tools, to each other, and to models. Anthropic's announcement is the primary source for what follows.
Graphic: the agent standards race at a glance. Sources: Anthropic; Linux Foundation. ByteHerald analysis.
What Anthropic gave away
A skill is deliberately simple: a folder containing a SKILL.md file that teaches an agent a specific task or domain, plus any scripts or reference material the work needs. Anthropic developed the format while building Claude Code, which handled general knowledge work well but lacked the procedural knowledge real business tasks require. Developers took to it fast: within weeks of the October 2025 launch, thousands of skills appeared on GitHub with loaders for Codex, Cursor, and OpenCode. Anthropic published Agent Skills as an open standard that December, and the numbers since are startling: more than 70 agent platforms have integrated it, including ChatGPT, Cursor, GitHub Copilot, and VS Code; about 1 million skills have been published; and installs approach 280 million. Those are Anthropic's figures, reported by the company rather than independently audited, and they should be read as such.
The donation is comprehensive: the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain all move to the Agentic AI Foundation (AAIF). Developers can keep using the standard and contributing to it without joining the foundation. Maintenance stays with engineers from Anthropic, Google, OpenAI, and other companies serving as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. The proposed roadmap covers versioned releases, a formal proposal process, and security guidance for people who write and install skills.
The stack, layer by layer
Six standards now sit under the AAIF roof, and they divide the agent economy into layers the way the internet's protocols divided networking. Here is the scoreboard:
The agent standards scoreboard: what each protocol governs and where it lives now. Sources: Anthropic, Oct 22, 2026; Linux Foundation announcements. ByteHerald compilation.
The two oldest entries do the heaviest lifting, and they are deliberately complementary. MCP is vertical: it standardizes how one agent reaches tools and data, a database, a calendar API, a file store. A2A is horizontal: it standardizes how independent agents find each other, delegate work, and exchange results without exposing their private reasoning. Google created A2A in April 2025, donated it to the Linux Foundation that June, shipped the stable v1.0 specification in March 2026, and moved it into the AAIF in August 2026, where it now sits alongside MCP with backing from more than 150 organizations, Forkast reported. Most products that expose an agent will eventually run both.
The four October 2026 additions fill the remaining layers. Open Responses, contributed by OpenAI, is a shared interface for talking to language models from different providers, so a request written once runs across vendors with little rewriting. OpenEnv, from Hugging Face, packages reinforcement-learning training environments as containers that speak both MCP and the Gym interface training libraries already use. OpenSandbox, from Alibaba, covers execution sandboxing, the contained environments where agents run untrusted code. Agent Skills sits above them all as the knowledge layer: the reusable procedures agents load on demand.
Twelve months to a neutral home
From a Claude Code experiment to ten Linux Foundation projects in twelve months. Sources: Anthropic; Linux Foundation. ByteHerald timeline.
The timeline is the story. In October 2025, Agent Skills was an Anthropic product feature. By December 2025 it was an open standard, the AAIF had launched with three projects, and MCP had its neutral home. In 2026, A2A stabilized and consolidated under the same roof. Now the skills standard follows, and the foundation jumps from three projects to ten in its first year, with founding membership growing from 49 organizations to more than 250. Whatever one thinks of the individual protocols, the institutional velocity is remarkable: the agent economy went from proprietary integrations to shared plumbing in about a year.
Why the labs keep giving away the rules
There is a strategy here, and it is older than AI. Donating a standard to a neutral foundation is the Linux playbook: commoditize the layer you do not monetize to grow the market for the layer you do. Anthropic does not sell protocols; it sells Claude. Google does not sell A2A; it sells Cloud. A shared, boring, interoperable plumbing layer means more agents get built, which means more inference gets bought. The donation is generosity and self-interest in the same envelope, and there is nothing wrong with that. Open standards with real multi-vendor maintenance beat corporate standards with press releases.
It is also worth naming the alternative the industry chose against. Much of the web's current response to AI agents is refusal: publishers, retailers, and banks blocking agent traffic outright, which we documented in our audit of the great agent lockout. Blocking and standardizing are the two competing answers to the same question, what do we do about software that acts on its own, and October 2026 advanced both. The blockers are winning the near term; the standardizers are playing for the decade.
The open questions
Three things will decide whether this stack matters or becomes committee ware. First, security: a million published skills with 280 million installs is a supply chain, and supply chains get attacked. The roadmap's promise of security guidance for skill authors and installers is the most important line in Anthropic's announcement, and the least detailed. We catalogued 2026's AI security incidents in Assume Breach, our security review; a backdoored skill package would fit that catalog's 2027 edition uncomfortably well.
Second, governance in practice. "Engineers serving as individuals" is the right structure, but the test comes with the first contentious proposal: a breaking change, a rejected security fix, a sponsor pushing its own extension. The AAIF's technical steering committees have not yet had their first real fight. Third, the missing layers: identity, payments, and liability for agent actions remain unstandardized, which is exactly where the lockout disputes are hottest. The plumbing is standardizing faster than the law, and that gap is where the next conflicts will live.
The test of the stack
Standards are promises; the test is what survives contact with competing interests. The first real signal will be Agent Skills' first formal spec version under the AAIF: multi-vendor maintenance working, or Anthropic remaining the de facto editor. The second is security: the 280-million-install figure makes the skills registry a target worth attacking, and the first backdoored skill package will show whether the promised guidance was real. The third is procurement language. When enterprise RFPs start requiring AAIF-standard agent interoperability, the race will have been won not by the best protocol but by the most boring one, which is how infrastructure victories always look.
By ByteHerald Staff. Anthropic's Agent Skills donation to the Agentic AI Foundation, from Anthropic's official announcement. Google's A2A protocol joining AAIF, reported by Forkast. Agent Skills registry scale (280-million-install figure) from Anthropic's announcement materials. Enforcement-layer context from our audit of the great agent lockout and our Assume Breach security review.
Comments
Post a Comment