Agentic AI Foundation Adds Four Open-Source Agent Projects

The Agentic AI Foundation, a Linux Foundation initiative launched in December 2025, added four open-source agent projects donated by Anthropic, OpenAI, Hugging Face and Alibaba at an October 22 event, bringing its hosted total to 10. The additions are Anthropic's Agent Skills, an open format for packaging reusable agent capabilities that already counts 70+ supporting platforms, over one million published skills and 280 million installs of the skills.sh registry in its first seven months; OpenAI's Open Responses, a multi-provider LLM interface specification with a compliance test suite; Hugging Face's OpenEnv, a containerized reinforcement learning environment framework compatible with MCP and Gym/Gymnasium; and Alibaba's Apache 2.0-licensed OpenSandbox for isolated execution. The foundation now has 170+ member organizations with platinum backing from AWS, Google, Microsoft and OpenAI; CTO Manik Surtani said "When one company holds the keys to a project, open is decided one release at a time.

Four of the biggest names in AI — Anthropic, OpenAI, Hugging Face, and Alibaba — walked into a Linux Foundation event on October 22 and donated the building blocks of the modern agent stack. The Agentic AI Foundation now hosts 10 open-source projects. This week’s four additions cover the layers developers have been patching together with glue code for the past year: reusable skills, a unified model interface, repeatable training environments, and sandboxed execution. What Is the Agentic AI Foundation? The AAIF launched in December 2025 under the Linux Foundation with three projects: Anthropic’s MCP, Block’s Goose, and OpenAI’s AGENTS.md. Ten months later it has 170+ member organizations, platinum backing from AWS, Google, Microsoft, and OpenAI, and now 10 hosted projects. The mission is neutral governance for agentic AI infrastructure — the idea being that if the foundation layer is open and vendor-neutral, no single company can hold the entire agent ecosystem hostage. When one company holds the keys to a project, open is decided one release at a time. Manik Surtani, CTO, Agentic AI Foundation The Four New Projects Agent Skills Anthropic Agent Skills is an open format for packaging reusable agent capabilities — instructions, scripts, templates, and resources bundled in a standard folder with a required SKILL.md file. Think of it as npm for agent capabilities. A contract-review skill can contain legal-approved procedures; a data-analysis skill can carry pre-approved query patterns. Skills are versioned artifacts that travel with applications rather than staying locked in vendor configurations. Anthropic donated the project to ensure “anyone building with skills can shape what’s next,” according to MCP co-creator David Soria Parra. The numbers suggest the format already found product-market fit before this announcement: 70+ agent platforms support it, over one million skills are published, and the skills.sh community registry https://github.com/anthropics/skills hit 280 million installs in its first seven months. Open Responses OpenAI Open Responses is a multi-provider LLM interface specification. It defines a shared request-response schema, consistent streaming events, and extensible tool-call patterns — with a compliance test suite to verify conformance. The goal is that developers write integration logic once and deploy it across model providers without rewriting anything when they switch. Databricks Model Serving and Vercel AI Gateway already conform to the spec, supporting Claude, Gemini, and Databricks-hosted models under the same interface. One caveat worth naming: Open Responses standardizes the interface, not the outputs. Models will still behave differently. But eliminating the adapter layer between them is a real productivity win. OpenEnv Hugging Face OpenEnv provides a standardized framework for creating and sharing reinforcement learning environments for agent training and evaluation. It packages environments as containerized services compatible with MCP and the Gym/Gymnasium training protocol — meaning an environment built for one agent harness can be dropped into another without modification. A benchmark result from the Turing Calendar Gym illustrates the case for well-defined interfaces: agents using explicit identifiers achieved roughly 90% task success, compared to around 40% with natural-language descriptions alone. OpenEnv also supports converting production failures into replayable test cases — effectively giving agents regression tests. OpenSandbox Alibaba OpenSandbox is a general-purpose sandbox platform for AI applications. It provides APIs for creating isolated environments, executing commands, and managing files — with Docker and Kubernetes runtimes, per-task isolation filesystem, resource limits, network policies , and multi-language SDKs. Licensed Apache 2.0. Alibaba’s contribution carries a clear governance philosophy: isolated execution is the safety primitive for agents that touch real systems — code execution, file writes, external API calls. OpenSandbox makes that primitive portable and open rather than locked behind a proprietary runtime. Why Now, and Why It Matters The fragmentation problem in agentic AI is well-documented and getting worse. Every framework ships its own tool-calling format. Every LLM vendor uses different streaming conventions. Building against one provider’s agent API creates the same lock-in developers spent a decade escaping in cloud infrastructure. According to the State of Agent Interoperability 2026 report https://openstandardagents.org/research/state-of-agent-interoperability/ , 85% of engineering leaders require vendor-neutral standards before committing agents to production, and enterprises currently spend roughly 40% of agent budgets on integration glue. The four new AAIF projects address those layers directly. Together they form a composable stack: skills define what an agent knows how to do, Open Responses defines how it talks to models, OpenEnv defines how it trains and evaluates, and OpenSandbox defines where it safely executes. None of these require each other — you can adopt one without the others — but used together they reduce the surface area of proprietary dependencies significantly. The Honest Assessment Standards without enforcement are aspirational. The AAIF provides formats and specifications, but whether a skill is actually portable across 70+ platforms, whether a provider genuinely conforms to Open Responses, whether OpenSandbox actually prevents credential leaks — those are questions developers have to verify in their own stacks. The foundation gave you the building blocks. Assembly is still your job. That said, the trajectory here is encouraging. MCP went from Anthropic internal tool to 110 million monthly SDK downloads and 10,000+ published servers. AGENTS.md reached 60,000 open-source projects. When AAIF standards get traction, they get real traction. Watching whether Open Responses achieves similar adoption as MCP in the next six months will tell you whether the foundation model actually works for interface standards. The Agent Skills repo is at github.com/anthropics/skills https://github.com/anthropics/skills . The full announcement is on the Linux Foundation press page https://www.linuxfoundation.org/press/agentic-ai-foundation-announces-four-new-projects-agent-skills-open-responses-openenv-and-opensandbox . Technical analysis of all five interoperability validation criteria is available from The Agentic Stack https://theagenticstack.substack.com/p/the-agent-stack-just-got-four-new .