The Agentic AI Foundation, the Linux Foundation group that already looks after the Model Context Protocol, has taken in four more projects from four rival companies. Anthropic gave it Agent Skills. OpenAI gave it Open Responses. Hugging Face gave it OpenEnv, and Alibaba gave it OpenSandbox. The foundation announced the move on 22 October in San Jose, at its AGNTCon + MCPCon North America event.
If you build with AI agents, or plan to, this is one of the more practical pieces of news this month. Each project solves a specific problem that every agent builder runs into, and moving them under a neutral foundation makes it more likely that they will work the same way across different companies' tools.
What the Agentic AI Foundation is
The Agentic AI Foundation (AAIF) launched in December 2025 with three projects: the Model Context Protocol (MCP), the goose agent framework and AGENTS.md, a simple file format that tells coding agents how to work in a repository. In its first year it has grown to 10 hosted projects, including Agent2Agent (A2A), Agent Router and agentgateway.
The idea is that the plumbing agents depend on should not belong to one company. "A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive," said Manik Surtani, the foundation's chief technology officer and co-founder.
The four new projects, in plain terms
1. Agent Skills (from Anthropic)
Agent Skills is an open format for packaging know-how. A skill bundles instructions, scripts and resources into a folder that an agent can load only when it needs it. A company could write one skill for "how we format invoices" or "how to run our test suite", and any agent that supports the format could pick it up.
"Without the right context, even the most intelligent agents fall short," said David Soria Parra, co-creator and lead maintainer of MCP at Anthropic.
Why it matters to you: you write a skill once instead of rewriting the same long prompt for every tool your team uses.
2. Open Responses (from OpenAI)
Open Responses is a specification for talking to large language models through one shared interface. It is based on OpenAI's Responses API and defines a common schema, streaming events and a way to plug in tools.
"Open Responses is thrilled to be joining the AAIF to further our mission of helping developers build great products," said Steve Coffey, a member of technical staff at OpenAI.
Why it matters to you: if model providers adopt it, switching from one model to another, or using several at once, should take fewer code changes.
3. OpenEnv (from Hugging Face)
OpenEnv is a framework and community hub for reinforcement learning environments, the practice worlds where agents are trained and tested. The environments run as containerised services and can be reached through MCP and the Gym interface that many researchers already use.
"Reinforcement Learning environments are a fundamental part of that stack," said Ben Burtenshaw, a machine learning engineer at Hugging Face.
Why it matters to you: teams that want to train or evaluate an agent on their own tasks get a shared way to package those tasks instead of building a test harness from scratch.
4. OpenSandbox (from Alibaba)
OpenSandbox is a platform for running the code an AI writes in an isolated space. It offers software kits in several languages, one set of sandbox APIs, Docker and Kubernetes runtimes, and a protocol for starting, running and stopping sandboxes.
"While OpenSandbox was created at Alibaba, we don't want its future to be defined by a single company," said Xiaobin Xu, head of development infrastructure at Alibaba.
Why it matters to you: letting an agent run code on your own servers is risky. A standard sandbox makes it easier to contain mistakes and attacks.
How the pieces fit together
Put the four projects next to the ones the Agentic AI Foundation already hosts and a rough stack appears:
- Talking to models: Open Responses
- Connecting to tools and data: MCP
- Knowing how to do a job: Agent Skills and AGENTS.md
- Talking to other agents: Agent2Agent
- Running code safely: OpenSandbox
- Training and testing: OpenEnv
None of these layers is new on its own. What is new is that the main pieces now sit in one neutral home, contributed by companies that compete hard everywhere else.
What to watch before you build on them
Handing a project to a foundation does not make it a standard overnight. A few things will decide whether these four stick:
- Adoption by rivals. Open Responses matters only if providers other than OpenAI support it. Agent Skills matters only if tools other than Anthropic's read the format. The press release does not list adoption numbers.
- Governance. Neutral governance is the point. Watch whether maintainers from outside the founding company join each project.
- Security. Skills can include scripts, and sandboxes run untrusted code. Both need careful review before you use them on real data, a point made clear by recent agent incidents we have covered, such as how 1Password is handling agent logins.
What it means for developers in Africa
For smaller teams, including many startups across Africa that build on top of several model providers to manage cost, shared standards are good news. They lower the price of switching between OpenAI, Anthropic, open-weight models and others as prices change. They also make it easier to hire, because skills learned on one tool carry over to the next.
The practical step for now: if you already use MCP, read the Agent Skills format and the Open Responses specification on the foundation's site and check which of your current tools support them. If you run agents that execute code, compare OpenSandbox with what you use today. These projects are open source, so trying them costs only time.
