Four Rivals Handed Their Agent Standards to a Neutral Foundation: Why That Helps You
Direct answer
The Agentic AI Foundation (AAIF), under the Linux Foundation, announced on October 22, 2026 at AGNTCon + MCPCon North America (San Jose) four new projects: Agent Skills (contributed by Anthropic), Open Responses (OpenAI), OpenEnv (Hugging Face) and OpenSandbox (Alibaba). With them, the foundation grows from 3 projects at launch (December 2025) to 10 in its first year, under open, vendor-neutral governance. The idea is simple and powerful: the rules by which agents gain skills, talk to models, train and run safely should not belong to a single company.
The four projects, layer by layer
- Agent Skills (Anthropic): how agents gain new, reusable capabilities. Anthropic reports the format is already adopted by 70+ platforms, with about 1 million skills and around 280 million installs, according to the week's announcement summary.
- Open Responses (OpenAI): how applications talk to models across providers, an interoperability layer so you do not rewrite your integration every time you change models.
- OpenEnv (Hugging Face): environments for training and evaluating agents, the part that lets you measure whether an agent actually does its job well before putting it in front of customers.
- OpenSandbox (Alibaba): where agents run safely, the isolated execution layer that separates the agent from your critical systems.
What already lived in the foundation
The four join projects already under the AAIF: the Model Context Protocol (MCP), the Agent2Agent (A2A) protocol, Agent Router, agentgateway, AGENTS.md and goose. The message from CTO and co-founder Manik Surtani sums up the bet: the community needs one place to build shared standards, with no single company controlling the outcome.
Why this matters to a business that does not write code
- Less lock-in: if every vendor invents its own way to connect agents to tools, switching vendors means starting over. Shared standards lower that switching cost.
- Portable skills: an agent that learns a skill on one platform should be able to use it on another. Agent Skills targets exactly that.
- Evaluation before production: OpenEnv targets the question every business owner should ask before deploying an agent: how do I know it works, and how do I prove it?
- Safe execution by design: OpenSandbox, and the sandbox conversation (Google announced its Agent Compute Engine the same week), shows the industry now treats isolated execution as infrastructure, not an extra.
The honest note
- The foundation was not born this week: a common headline error is to say the AAIF launched on October 22. The foundation started in December 2025 with three projects; what is new are the four projects.
- Open standard does not mean adopted standard: four companies handing technology to neutral governance is a strong signal, but real adoption is measured in implementations, not announcements. If your vendor promises compatibility with these projects, ask for the concrete implementation and date.
- Neutral governance does not end competition: Anthropic, OpenAI, Hugging Face and Alibaba still compete for your contracts. What changes is that they compete on shared rules, and that leaves you with more negotiating power.
Related links
- How an AI agent works inside (without jargon)
- What an AI agent is and how a real business uses one
- How to connect your CRM with AI agents (practical guide)
Frequently asked questions
What did the Agentic AI Foundation announce on October 22, 2026?
Four new projects under its governance: Agent Skills (Anthropic), Open Responses (OpenAI), OpenEnv (Hugging Face) and OpenSandbox (Alibaba). The foundation grows from 3 to 10 projects in its first year.
Is the AAIF a new foundation?
No. It launched in December 2025 with three projects, under the Linux Foundation. What was announced on October 22 are the four new projects, not the foundation.
What does each new project cover?
Agent Skills: how agents gain capabilities. Open Responses: interoperability between applications and models. OpenEnv: agent training and evaluation. OpenSandbox: safe agent execution.
How does an open agent standard help me?
It reduces lock-in to a single vendor, makes your agents' skills portable, gives you ways to evaluate them before production, and treats safe execution as an infrastructure layer.
