Four frontier labs walked onto the AGNTCon stage in San Jose on October 22 and handed the layers of the agent serving stack to someone else. Anthropic donated Agent Skills. OpenAI donated Open Responses, derived from its Responses API. Hugging Face contributed OpenEnv, a containerized RL environment framework. Alibaba contributed OpenSandbox, a multi-language SDK sandbox. All four landed inside the Linux Foundation's Agentic AI Foundation (AAIF), which now hosts ten projects after launching with three in December 2025.

The donations fill in the layers that were previously proprietary at the top of the stack: the capability format, the model-serving interface, the training environments, and the execution sandbox. Governance of each now sits with a vendor-neutral body rather than the labs that built them.

The Agent Skills numbers explain why this move was functionally inevitable. Anthropic launched the format in October 2025, published it as an open standard that December, and watched it propagate: more than 70 agent platforms have integrated it, roughly a million skills have been published, and installs are approaching 280 million. ChatGPT, Cursor, GitHub Copilot, and VS Code already read the same SKILL.md files Claude Code does. Community loaders like Codex and OpenCode fill the gaps. When a format reaches that distribution, keeping the governance at one lab is less a strategic asset than a liability.

"Without the right context, even the most intelligent agents fall short," said David Soria Parra, Co-Creator and Lead Maintainer of MCP at Anthropic, in the AAIF announcement.

Steve Coffey, Member of Technical Staff at OpenAI, framed the Open Responses contribution as a continuation of the same logic: a shared interface for how models receive and return tool calls, no longer bound to any single provider's SDK.

Manik Surtani, CTO and co-founder of the AAIF, is the operator on the receiving side. Less than a year ago his foundation accepted Anthropic's donation of the Model Context Protocol. The pattern is now legible: labs ship a protocol, let it propagate, then transfer stewardship once the network effects are irreversible. It's the governance arc that TCP/IP took to the IETF in the late 1980s, compressed to about eighteen months.

There's also a quiet security story. The Agent Forge scanned 9,930 public SKILL.md files across 277 GitHub repositories and flagged 30, about 0.33%: 14 for piping downloads into a shell, 8 for reading entire key files to pull one variable. Hand review of 26 of them found real patterns and no malice. The scan surfaced no malware. For a format defined only by a name field of 1–64 characters and a description up to 1024, that's a usefully boring result.

Boring is the point. The interesting layers of the stack are no longer the ones any single lab controls.

Sources

https://www.linuxfoundation.org/press/agentic-ai-foundation-announces-four-new-projects-agent-skills-open-responses-openenv-and-opensandbox
https://aaif.io/news/agentic-ai-foundation-announces-four-new-projects
http://www.prnewswire.com/news-releases/agentic-ai-foundation-announces-four-new-projects-agent-skills-open-responses-openenv-and-opensandbox-302916689.html
https://www.aichatdaily.com/ai-news/agentic-ai-foundation-adds-anthropic-s-agent-skills
https://www.developersdigest.tech/blog/agent-skills-agentic-ai-foundation-what-changes-2026