The Agentic AI Foundation added four projects on 22 October, extending its remit from protocols into reusable skills, model interfaces, training environments and execution infrastructure. Anthropic is donating Agent Skills, while OpenAI, Hugging Face and Alibaba contribute projects at other layers of an agent system. The change concerns who steers shared infrastructure, rather than a new model release.

Based on 4 reviewed sources · See sources

The format gets a different steward

Anthropic’s donation includes the Agent Skills name and logo, its GitHub organisation and the agentskills.io domain. The company says people can keep using and contributing to the standard without becoming AAIF members. Engineers from several companies will maintain it as individuals, with David Soria Parra and Mahesh Murag as lead maintainers.

A skill packages procedural instructions in a SKILL.md file, with optional scripts and supporting resources. Anthropic says more than 70 agent platforms have integrated the format. That is its reported integration count, rather than evidence that every skill runs identically across those products. Its original announcement below explains the transfer.

From the source · ClaudeDevs ·

Anthropic announces the Agent Skills donation. Platform integration figures come from the company; the follow-up thread describes governance and a proposed roadmap.

View the original post on X ↗

Four projects, four different jobs

The foundation’s announcement places the additions alongside existing projects including MCP, A2A, goose and AGENTS.md. It now hosts ten projects. The four new arrivals address different parts of the stack; joining the same foundation does not turn them into one product.

The detail

What each new project covers

AAIF announcement · 22 October 2026

What each new project covers. AAIF announcement · 22 October 2026
Project / contributorRole
Agent Skills · AnthropicPackages instructions and resources that agents load as needed.
Open Responses · OpenAIDefines a shared request, output and streaming interface across model providers.
OpenEnv · Hugging FaceProvides interoperable environments for agent training and evaluation.
OpenSandbox · AlibabaProvides sandbox APIs and runtimes for executing agent workloads.
These descriptions explain project scope. They do not establish interoperability test results, security guarantees or availability in a particular hosted product.Source: Agentic AI Foundation project announcement · 2026-10-23

A shared folder is a starting point

The specification includes an optional compatibility field for environment requirements. Its allowed-tools field remains experimental, with support varying between implementations. For teams packaging a repeatable research or publishing workflow, those details determine how much travels with the instructions.

Anthropic’s proposed roadmap covers versioned releases, a formal proposal process and security guidance. Those are next steps, rather than newly delivered safeguards. Neutral stewardship gives more parties a route to shape the format; the practical test remains whether a particular skill works with the tools and permissions of its destination.

Sources

Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.