Agentic AI Foundation Adds Agent Skills and Open Responses

On 22 October, the Agentic AI Foundation (AAIF) announced four new projects at AGNTCon + MCPCon North America in San Jose. Two of them came from competing AI model companies. Anthropic contributed Agent Skills, and OpenAI contributed Open Responses. Both now sit under the same vendor-neutral roof as the Model Context Protocol (MCP).

That’s the news. The more useful question is what each project does, and why a company would give it away.

The four projects

Project	Contributed by	What it does
Agent Skills	Anthropic	A format for packaging instructions, scripts and reference files that an agent loads when it needs them
Open Responses	OpenAI	An open spec for one request format that works across model providers, based on OpenAI’s Responses API
OpenEnv	Hugging Face	A framework for building reinforcement learning environments, with a standard agent interface and the Gym training protocol
OpenSandbox	Alibaba	A sandbox platform for running agent code, with SDKs and Docker and Kubernetes runtimes

The foundation says the four cover different layers of the agent stack: what agents can do, how apps talk to models, how agents get trained and evaluated, and where agents run. AAIF now hosts 10 projects. It launched in December 2025 with three.

Why Agent Skills matters to builders

A skill is a folder with a SKILL.md file and whatever supporting files it needs, according to the secondary report. The report says an agent loads these files only when it needs them.

According to the same secondary report, Anthropic built the format while working on Claude Code, and published it as an open standard in December 2025. It also says more than 70 platforms have integrated it. We couldn’t check that count against a primary source, so treat it as a claim.

My read: what changes now is that the spec no longer belongs to one company alone.

Why neutral governance is the real story

My read: a project with one owner can change its rules whenever that owner wants. A foundation makes that harder, because changes need a wider group to agree. For a standard many tools depend on, that is the point. The announcement itself doesn’t describe the voting process.

Manik Surtani, CTO and co-founder of AAIF, says the foundation exists to give the community “one place to build shared standards, protocols, and infrastructure, with no single company controlling the outcome in a way that puts the community first.” My read: that is a promise about process. Whether it holds depends on who makes the decisions and how they are made, and the announcement doesn’t say.

My read: Anthropic and OpenAI both gain here. Rival tools may find it easier to adopt Agent Skills, since the format is no longer one company’s product, and Open Responses is meant to let developers run one request across providers with minimal translation, so switching is cheaper. The cost for both is control. They give up the power to change the format alone, which is the price of having others build on it.

What to watch

My read on adoption: a foundation home is a start. The test is whether other agent products read the same skill folders and accept the same request format.
My read on governance: neutral in name only is a real risk. The membership list and governance rules will show how neutral it is.
My read on overlap: the foundation already hosts A2A, AGENTS.md and goose, so four new projects can also mean four new things to learn.

Related reading: our explainer on adding an MCP server to Claude covers the protocol that AAIF already governed before this week.

Sources

Note on sources: The first two are the foundation’s own release and are the primary sources for what each project is and who contributed it. The adoption count, the history of Agent Skills and the SKILL.md layout come from the secondary outlet, not the release. I did not open the Agent Skills spec repository, so the layout is not checked against the spec itself.