Anthropic donated Agent Skills, its format for packaging capabilities that AI agents can load when needed, to the Agentic AI Foundation (AAIF) on 22 October 2026. The foundation announced the donation at AGNTCon + MCPCon North America in San Jose, alongside projects contributed by OpenAI, Hugging Face and Alibaba, in its release. The move places governance of the Agent Skills standard with a foundation under the Linux Foundation rather than with Anthropic alone.
Key points
- Anthropic has contributed Agent Skills to the Agentic AI Foundation, which says it now hosts 10 projects.
- The format packages instructions, scripts and other resources for agents to load on demand.
- OpenAI, Hugging Face and Alibaba contributed projects covering model interfaces, training environments and places where agents run.
AAIF grows to 10 hosted projects
AAIF says it now hosts 10 projects, compared with three when it launched in December 2025. Its existing projects include the Model Context Protocol (MCP), which Anthropic previously donated, as well as Agent2Agent, Agent Router, agentgateway, AGENTS.md and goose. Anthropic co-founded the foundation with Block and OpenAI, Simon Willison reported. The projects are governed under the foundation’s stated vendor-neutral approach.
Manik Surtani, AAIF’s chief technology officer and co-founder, said the four new contributions give developers a shared place to work on standards and infrastructure without one company controlling their direction. For companies building agents, that distinction concerns who can help shape the formats their systems depend on. The foundation describes the new projects as covering four different parts of an agent system, from acquiring a capability to running code in a sandbox.
Anthropic puts Agent Skills under shared governance
Agent Skills gives an agent a way to take up specialised instructions, scripts and resources as a reusable package. Think of a folder containing directions and the materials needed for a particular task: the agent can open it when that task arises. The foundation says the format is intended to let that knowledge travel between compatible agent products. David Soria Parra, Anthropic’s co-creator and lead maintainer of MCP, said the company launched Agent Skills about a year before the donation and wants people building with it to participate in its development.
Writing a report could draw on a packaged set of instructions and resources when the task calls for them. That would depend on the agent supporting the format and on a suitable skill being available, but the package could then be loaded without being tied to a single agent product.
The foundation says more than 70 agent platforms support Agent Skills, more than 1 million skills have been published and the format has attracted nearly 280 million installations, according to Unite.AI. Those are the foundation’s adoption figures. They count platforms, published skills and installations, rather than the performance of an agent carrying out a task. The foundation also says the specification is available under Apache 2.0 and CC BY 4.0 licences.
Under AAIF, the foundation says, Agent Skills will have a technical steering committee whose core maintainers come from multiple organisations and serve as individuals. That arrangement gives developers from different companies a route into decisions about the format. It also makes the governance change more specific than a promise to keep the specification open: maintainers outside Anthropic can participate in directing the project.
OpenAI, Hugging Face and Alibaba contribute projects
OpenAI contributed Open Responses, a specification based on its Responses API for exchanging requests and results between applications and models from different providers. AAIF says it sets out a common structure for those exchanges, including streamed output and calls to other tools. A common structure could reduce the amount of provider-specific work an application needs, while still allowing providers to identify their own extensions separately.
The foundation identifies Databricks Model Serving and Vercel AI Gateway as existing implementations, Unite.AI reported. Databricks translates requests from an Open Responses endpoint into the formats used by the models it serves. That translation is a concrete part of making a common interface work: the shared request format sits in front of provider-specific ones, without requiring a model to be retrained.
Hugging Face contributed OpenEnv, which AAIF describes as a framework for environments used to train and evaluate agents through reinforcement learning. An agent can start an episode, take an action and receive an observation and a numerical reward through its interface. Unite.AI reported that a team at Turing using an OpenEnv-based calendar environment found roughly 90% task success when requests supplied explicit calendar identifiers, against roughly 40% when requests instead used natural-language descriptions. That comparison concerns those two ways of phrasing tasks in that environment.
Alibaba contributed OpenSandbox, which AAIF describes as a platform for running AI applications in sandbox environments. Its protocol defines how those environments are started and used, while its software supports Docker and Kubernetes runtimes. Xiaobin Xu, Alibaba’s head of development infrastructure, said bringing the project to AAIF would allow developers and organisations beyond Alibaba to help shape its future. The four donations were announced during AGNTCon + MCPCon North America, held on 22 and 23 October at the San Jose McEnery Convention Center.
