Anthropic and OpenAI Donate Agent Standards to Linux Foundation

On Thursday, October 22, Anthropic transferred the Agent Skills standard to the Agentic AI Foundation (AAIF), an entity hosted by the Linux Foundation. In the same announcement, the foundation received Open Responses, a specification derived from OpenAI's Responses API, as well as OpenEnv from Hugging Face, and OpenSandbox from Alibaba. With these four additions, AAIF now hosts ten projects in its first year, according to the foundation's statement.

The statistics behind Agent Skills explain the significance of the donation. According to Anthropic, more than 70 agent platforms have integrated the standard since its release in December 2025, including ChatGPT, Cursor, GitHub Copilot, and VS Code. The company claims that 1 million skills have been published and nearly 280 million installed.

Changes in Governance

A skill is a folder containing an SKILL.md file that teaches an agent to perform a task or master a subject, accompanied by scripts and reference material. Anthropic hands over the name and logo of the standard, the organization on GitHub, and the domain agentskills.io to AAIF. Engineers from Anthropic, Google, OpenAI, and other companies will maintain the specification as individuals, with David Soria Parra and Mahesh Murag as the primary maintainers.

This follows the same path as the Model Context Protocol (MCP), which Anthropic donated to AAIF in December 2025, when it founded the entity with Block and OpenAI. With this donation, the company relinquishes sole control over the technical roadmap of the standard. The initial roadmap proposal includes numbered versions, a formal proposal process, and security guidelines for those writing and installing skills.

The inclusion of Alibaba in the same batch carries weight. OpenSandbox provides an isolated execution environment for agents, supporting Docker and Kubernetes, and places a Chinese company as a direct contributor to a set of standards dominated by leading American labs. Among the platinum-level members since its launch, the foundation includes Amazon Web Services, Google, Microsoft, OpenAI, Anthropic, Block, Bloomberg, and Cloudflare.

Impact on the CIO Budget

For buyers of agents, the change mitigates a concrete risk: writing internal processes in a format controlled by a single vendor. A skill written for Claude Code should theoretically run on any of the more than 70 platforms that adopted the standard. Open Responses makes an equivalent move on OpenAI's side by proposing a common interface for calls to models from different providers.

The fragile aspect lies in the impressive number itself. Nearly 280 million installations of packages that may contain executable scripts form a supply chain surface comparable to that of repositories like npm and PyPI, which have lived with malicious packages for years. Anthropic itself included security guidelines in the initial roadmap, indicating that these do not exist in a standardized form today. Until they do, the decision about which skills a corporate agent can carry remains with each company's security team.

There is also a limit to the neutrality argument. An open standard does not mean equal execution: the performance of a skill depends on the model interpreting it, and the commercial competition among labs has shifted from the interface to the quality of the agent.

Market Readings

In the United States, where the three founders of AAIF are based, the donation solidifies a governance model that has already worked with the MCP. Competitors contribute specifications to the same foundation and compete in the product.

In India, the effect reaches through integrators. TCS announced a partnership with Anthropic in June 2026 to bring Claude to regulated sectors, and Infosys secured a similar agreement to build agents in these markets. For these companies, a neutral skill format transforms the process knowledge they accumulate in banking and insurance clients into a reusable asset across platforms, no longer confined to a contract with a single lab.

In China, Alibaba's entry indicates that at least part of the local industry prefers to participate in the standard rather than build a parallel. It is the only one of the four contributions in the batch that does not come from an American or Franco-American company, and it joins a foundation where all eight platinum members mentioned above are from the United States.

For consulting firms that build agents for clients in multiple countries, the practical point is different. If the skills standard establishes itself as the MCP did for connectors, the sellable asset will shift from integration with a specific model to the library of documented processes, which the client can take with them when switching vendors.