SOLUTIONSQuick answer
On 22 October 2026 the Agentic AI Foundation, a Linux Foundation project, took on four open AI agent projects: Anthropic's Agent Skills, OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox. Together with MCP and AGENTS.md they form a vendor-neutral stack for building AI agents, which reduces lock-in to any single model provider.
Key takeaways
- The Agentic AI Foundation (AAIF) now hosts 10 projects, up from 3 at its December 2025 launch.
- Agent Skills packages instructions, scripts and resources that agents load on demand, so know-how works across agent products.
- Open Responses is a shared request and streaming format, based on OpenAI's Responses API, for calling different model providers the same way.
- OpenEnv (training environments) and OpenSandbox (safe code execution) cover the testing and runtime side of agents.
- For businesses, open standards mean you can switch model providers without rebuilding your AI features.
What did the Agentic AI Foundation announce?
At AGNTCon + MCPCon North America in San Jose on 22 October 2026, the Agentic AI Foundation announced four new hosted projects. AAIF is part of the Linux Foundation and gives AI agent projects neutral, open governance, so no single company controls their direction. Its founding projects were the Model Context Protocol (MCP), goose and AGENTS.md; it also hosts A2A, Agent Router and agentgateway.
| Project | Contributed by | What it does |
|---|---|---|
| Agent Skills | Anthropic | An open format for packaging instructions, scripts and resources that agents load only when needed |
| Open Responses | OpenAI | An open specification for multi-provider LLM interfaces: shared schema, consistent streaming events, extensible tools |
| OpenEnv | Hugging Face | A framework and hub for reinforcement learning environments, exposed through an agent interface, MCP and the Gym protocol |
| OpenSandbox | Alibaba | A sandbox platform for running AI-generated code safely, with SDKs, unified APIs and Docker or Kubernetes runtimes |
These are existing projects moving into neutral governance, not brand-new products. The signal is that the largest AI companies are agreeing on shared building blocks rather than competing on incompatible ones.
What are Agent Skills?
A skill is a folder of know-how an agent can pick up when a task calls for it: written instructions, example files and helper scripts. Instead of stuffing every rule into one giant prompt, the agent reads a short description of each skill and loads the full skill only when it is relevant. That keeps context small and lets teams reuse the same skill, for example "how we write release notes" or "how to fill in our invoice template", across different agent tools.
What is Open Responses?
Open Responses is a specification for talking to language models. It is based on the OpenAI Responses API and defines a common schema for requests, tool calls and streaming events. If several providers implement it, an application can send the same request to different models with little translation code, which makes it easier to compare models, fall back when one is down, or switch on price.
How do MCP, Agent Skills and Open Responses fit together?
- Open Responses: how your app talks to the model.
- MCP (Model Context Protocol): how the model reaches your tools and data, such as a CRM, database or file store.
- Agent Skills: the reusable know-how the agent applies to a task.
- AGENTS.md: instructions for coding agents working in a code repository.
- OpenSandbox: where agent-written code runs without touching your real systems.
- OpenEnv: where agents are trained and tested against realistic tasks.
What does this mean for businesses building AI features?
- Build on standards, not one vendor's SDK: connect tools through MCP and keep model calls behind a thin layer, so changing provider is a configuration change.
- Capture your know-how as skills: company processes written once can be reused by every agent tool your team adopts.
- Sandbox anything an agent executes: agents that run code or browse should do it in isolation, with least-privilege credentials.
- Plan for security from day one: this month a single attacker reportedly used AI agents to breach 100+ websites (see our write-up). Exposed agent dashboards and leaked keys were part of that story.
Naarvix builds custom software and AI features for clients worldwide. If you are planning an AI assistant, an internal agent or an AI feature in your product, see our custom software development service or contact us.
Sources
Frequently asked questions
Who runs the Agentic AI Foundation?
It is a Linux Foundation project that provides open, vendor-neutral governance for AI agent projects such as the Model Context Protocol, AGENTS.md and goose. Contributing companies hand their projects to the foundation so the wider community can shape them.
Do I need Agent Skills to build an AI agent?
No. Skills are a convenient, portable way to package instructions and scripts. Small agents can work with a single prompt; skills help when an agent handles many kinds of tasks or several tools should share the same know-how.
Is Open Responses the same as the OpenAI API?
It is based on the OpenAI Responses API but published as an open specification, so other model providers and tools can implement the same request and streaming format.