The Agentic AI Foundation just took on four new open projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. If you build with AI agents, that means fewer one-off formats and less rewriting when you switch model providers. Nothing breaks today, but the plumbing under agent tools is getting standardized.
Here are the facts worth knowing before we get into what each project does.
- Announced October 22, 2026, at AGNTCon + MCPCon North America in San Jose.
- The foundation now hosts 10 projects, up from three at its launch last December.
- The four newcomers cover skills, model APIs, training environments and sandboxes.
- The foundation sits under the Linux Foundation, so no single company owns the projects.
What the Agentic AI Foundation actually added
The foundation started with three projects: the Model Context Protocol (MCP), goose and AGENTS.md. Per the Linux Foundation announcement, it also hosts Agent2Agent (A2A), Agent Router and agentgateway. The four new ones fill gaps higher and lower in the stack.
The table below shows who gave what, and what it does in plain terms.
| Project | Contributed by | What it does |
|---|---|---|
| Agent Skills | Anthropic | An open format for packaging instructions, scripts and resources that an agent loads only when it needs them. |
| Open Responses | OpenAI | A shared spec, based on the OpenAI Responses API, for talking to different model providers the same way. |
| OpenEnv | Hugging Face | A framework and hub for reinforcement learning environments that run as containerized services. |
| OpenSandbox | Alibaba | A sandbox platform with multi-language SDKs and Docker and Kubernetes runtimes for running agent code safely. |
Most readers will touch the first two directly. The last two matter more to teams that train models or run agents in locked-down environments.
Agent Skills: know-how you can carry between agents
A skill is a folder of instructions, scripts and reference files that an agent reads when a task calls for it. Think of it as a how-to binder the agent opens only when needed, instead of stuffing everything into one giant prompt.
Anthropic released the format last December, and SiliconANGLE notes that a common format makes skills easier to share across teams. Moving it to neutral ground is the real news. Other vendors can now shape the format without asking one company for permission.
If you’ve seen Google’s free Android Skills playbooks for coding agents, this is the same idea getting a proper home. A skill you write once should work in more than one tool.
Open Responses, OpenEnv and OpenSandbox in plain English
Open Responses tackles a boring but costly problem. Every provider has its own API, so switching models often means rewriting code. A shared schema and consistent streaming events let you write a request once and point it at different providers.
That fits a trend we covered in OpenAI models arriving on Google Cloud: models are becoming easier to swap, and the glue code is the part people want to standardize.
OpenEnv gives researchers a common way to build the practice worlds where agents learn. Hugging Face says the committee behind it includes contributors from Microsoft, Meta-PyTorch and Modal. OpenSandbox gives agents a controlled box to run code in, with the same APIs whether you use Docker or Kubernetes.
Why neutral ownership matters to you
Open formats owned by one company tend to drift toward that company’s products. A foundation setting doesn’t remove all politics, but it does spread control across rivals who all have a stake in compatibility.
Anthropic’s David Soria Parra said the foundation has done important work giving MCP a neutral home, and Agent Skills follows the same path. Alibaba’s Xiaobin Xu put it more bluntly: the future of OpenSandbox shouldn’t be defined by a single company.
There’s a caveat. The announcement doesn’t state licenses, and it doesn’t give setup steps for the new projects. Check each project’s repository before you build anything critical on it.
What to do this week
You don’t need to change anything today. A few small moves will put you ahead, though.
- If you write agent prompts or tools, look at how Agent Skills structures a skill folder before inventing your own layout.
- If you call several model APIs, read the Open Responses spec and see how much of your wrapper code it would replace.
- If you run agents that execute code, compare OpenSandbox’s runtimes with your current setup.
Each of these is a reading task, not a migration. The payoff comes later, when your tools and your vendors agree on the same formats.
Frequently Asked Questions
Who runs the Agentic AI Foundation?
It’s hosted by the Linux Foundation and launched in December 2025 with MCP, goose and AGENTS.md. It now hosts 10 projects, according to the announcement.
Is Agent Skills only for Claude?
No. It’s an open format for packaging agent instructions, scripts and resources, and the donation is meant to let anyone building with skills shape it.
Do I need to switch to Open Responses?
No. It’s a spec based on the OpenAI Responses API that lets you write a request once and run it across providers. It’s worth reading if you juggle several models.
Does this change anything on my phone or PC?
Not directly. These are developer standards, though apps you use may adopt them over time.
Bottom line
Standards rarely make headlines, but they decide which tools survive. Giving skills, model APIs, training environments and sandboxes a neutral home is a sensible bet, and developers who learn these formats early will have less to redo later. Watch for licenses and real adoption by rival vendors before you treat any of this as settled.
Leave a Reply