Anthropic Donates Agent Skills to the Agentic AI Foundation, Joining OpenAI and Hugging Face

The Agentic AI Foundation announced its second wave of projects on October 22, and the roster reads like a map of the agent industry’s fault lines agreeing to interoperate: Agent Skills, the open standard for giving agents specialized knowledge and workflows, donated by Anthropic; Open Responses, an LLM API spec, donated by OpenAI; OpenEnv from Hugging Face; and OpenSandbox. The AAIF launched on December 9, 2025 under the Linux Foundation, co-founded by Anthropic, OpenAI, and Block with the Model Context Protocol as its founding project, and has since grown to more than 70 participating organizations. Anthropic framed the donation on the project page as the natural next step after MCP.

What Agent Skills is, and why it was worth donating

Agent Skills is the format behind one of the more quietly effective ideas in agent engineering: instead of stuffing every possible instruction into a system prompt, skills are folders of instructions, scripts, and reference material that an agent loads only when a task requires them. The pattern solves a real context-economics problem, and it has spread fast: Anthropic reports the format has been integrated by more than 70 platforms. We have covered the practical side of this from the builder’s angle, how to create Agent Skills and how they differ from tool-configuration conventions like cursor rules; donation to a foundation changes it from an Anthropic feature into neutral infrastructure that any vendor can implement without permission.

The foundation as the industry’s treaty organization

The composition of this announcement deserves more attention than any single project. Ten months ago the AAIF was an unusual truce: two labs whose agents compete directly, plus Block, putting their protocols into a shared body. The second wave adds Hugging Face, and the projects now cover the full agent stack: MCP for tool connectivity, Agent Skills for knowledge and workflows, Open Responses for API compatibility, OpenEnv for environments, and OpenSandbox for execution. This is how the agent ecosystem is resolving the standards war that the cloud and model layers are still fighting; the same week federal supply-chain rules split model procurement along national lines, the tooling layer is consolidating around neutral governance. Vendors learned the lesson of containers and Kubernetes: the protocol layer is safe to share precisely because the layers above it are where competition pays.

What to watch

Three signals will tell you whether the AAIF becomes Linux-style infrastructure or a parking lot for press releases. First, implementation breadth beyond the founders: 70 participating organizations is the claim, and the test is whether non-foundation agent runtimes ship MCP and Skills support as a default rather than an integration option. Second, whether Open Responses gains traction against the de facto OpenAI-compatible API conventions that already saturate the market; a donated spec only matters if it consolidates behavior that is currently diverging. Third, governance behavior under conflict: the first time a foundation member ships a breaking change or a competing protocol, the neutrality claim gets tested. For builders, the practical takeaway today is simple: skills, MCP, and typed tool interfaces are becoming the portable resume of an agent’s capabilities, and writing against the open versions is no longer an idealism tax; it is the default path with the longest support runway.