Anthropic and OpenAI Hand Agent Tools to the Linux Foundation

Four of the companies racing to build AI agents spent Wednesday doing something unusual for rivals. They gave their work away. At AGNTCon and MCPCon in San Jose, the Agentic AI Foundation, a body set up under the Linux Foundation, said it had taken in four new projects donated by Anthropic, OpenAI, Hugging Face and Alibaba.

The projects address the plumbing that every agent builder has to lay. Anthropic contributed Agent Skills, a format for packaging reusable instructions and tools so an agent can pick up a new capability without being retrained. OpenAI handed over Open Responses, a specification for a shared model API that standardises the basic pieces, messages and tool calls, so developers no longer have to rewrite their code for each vendor. Hugging Face gave OpenEnv, which covers the training environments agents learn in, and Alibaba donated OpenSandbox, which handles isolated execution so an agent can run code without touching the host system.

Standards over lock-in

The pitch behind all four is interoperability. An agent written against one lab's tools has, until now, tended to stay there. Open specifications let the same skill or the same API call work across products from companies that otherwise compete hard for the same customers. Anthropic said it transferred the Agent Skills name, logo, GitHub organisation and the agentskills.io domain to the foundation, and that developers can keep using the standard, or contribute to it, without joining.

The adoption figures come from the labs themselves and should be read as their own claims rather than independent counts. Anthropic points to more than a million published skills and close to 280 million installs, and says over 70 agent platforms, including ChatGPT, Cursor and GitHub Copilot, already support the format.

A growing commons

The foundation now hosts ten projects. The newcomers sit alongside the Model Context Protocol, the Agent2Agent protocol, the agentgateway and the AGENTS.md convention, among others. It was founded in December 2025 by Anthropic, Block and OpenAI, and the fact that those same companies are still feeding it suggests the open-standards approach is holding rather than fracturing into competing camps.

There is a practical reason for the generosity. Agents become more useful as more of them can talk to each other and to the wider web, and no single lab benefits from a hundred incompatible dialects. We have watched the same companies push agents outward in recent weeks, from Anthropic putting one agent in charge of a thousand to Claude Duet working on its own and Walmart opening a door for shopping agents. Shared rails are what let any of that scale beyond one company's walled garden.