Agentic AI Foundation Adds OpenAI's Open Responses and Anthropic's Agent Skills

TL;DR

On October 22 the Agentic AI Foundation (AAIF), the Linux Foundation home of MCP, announced four new projects at AGNTCon + MCPCon North America in San Jose. OpenAI gave it Open Responses, a vendor-neutral spec based on its own Responses API. Anthropic gave it Agent Skills, the SKILL.md folder format that more than 70 agent platforms already load. Hugging Face added OpenEnv for RL training environments, and Alibaba added OpenSandbox for running agent code in isolation. The foundation now hosts 10 projects, up from three when it launched in December 2025. If you build agents, the API shape, the skill format, the training harness and the sandbox now all sit under one neutral governance model.

What moved, and who moved it

The AAIF press release sorts the four donations by layer of the agent stack, one each for capabilities, model interoperability, training and evaluation, and secure execution. Each layer came from a different company.

The foundation's full roster is now the Agent2Agent (A2A) protocol, Agent Router, agentgateway, AGENTS.md, goose, the Model Context Protocol, and the four newcomers. "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 CTO and co-founder.

Simon Willison wrote in December 2025 that he "wouldn't be surprised" to see Agent Skills end up at AAIF. Ten months later he gets to link back to himself, the rarest pleasure in tech blogging.

Open Responses: OpenAI hands over the shape of its API

Open Responses launched on January 15 as an open spec modeled on OpenAI's Responses API, with Hugging Face, Vercel and OpenRouter among the early adopters. Per the AAIF write-up, it defines a common format for requests and responses: sending context to a model, describing tools, receiving tool calls, returning tool results and streaming output. It only covers the API layer, so a provider can adopt it without retraining anything, and tool execution and permissions stay with your agent.

Think of it as a plug standard. It fixes the socket and leaves the power station alone. Your client code talks to one request shape, and whatever model sits behind the endpoint still behaves like itself. Extensions get provider-specific names, so one vendor's extra streaming events (timing, cache data) don't collide with another's, and agents that don't recognize them can ignore them.

The AAIF post points to two implementations. Databricks Model Serving exposes an Open Responses endpoint for Claude, Gemini and Databricks-hosted open models, translating each request into the provider's native format. Vercel AI Gateway exposes an OpenResponses API too. The project site's community wall lists NVIDIA, Vercel, OpenRouter, Hugging Face, LM Studio, Databricks, Red Hat, AWS, Ollama, OpenAI, vLLM and Llama Stack.

Anthropic and Google are not on that list. You can reach Claude and Gemini through Open Responses on Databricks because Databricks translates, and neither company appears as an implementer of its own. A neutral home makes it easier for them to sign up later, but nobody has to.

Agent Skills: the format already won, now it changes owners

A skill is a folder holding a SKILL.md file plus any scripts or reference material the agent needs, loaded on demand. Anthropic says more than 70 agent platforms have integrated the format since it was published in December 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code. "So far, 1 million skills have been published and nearly 280 million installed."

Anthropic is donating the Agent Skills name and logo, the GitHub organization and the agentskills.io domain. Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals, with David Soria Parra (co-creator of MCP) and Mahesh Murag as lead maintainers. You don't need AAIF membership to use the format or contribute to it.

The proposed roadmap covers versioned releases, a formal proposal process and security guidance for people who write and install skills. That last item is still only a proposal. A skill can ship scripts, and 280 million installs is a large surface for a format that has no versioned releases yet.

OpenEnv and OpenSandbox: the less famous half

OpenEnv

OpenEnv splits RL training into a client that runs episodes and an environment server that defines the task. Clients call reset(), step() and state() and get back an observation, a numeric reward, a done flag and metadata. Environments are packaged with Docker and published to Hugging Face Spaces. One example from the AAIF post: Turing's Calendar Gym saw roughly 90% task success when tasks named explicit calendar identifiers, and roughly 40% when the same tasks used natural-language descriptions. That gap is a decent argument for testing agents on the messy version of a task.

OpenSandbox

OpenSandbox is an Apache 2.0 isolation platform with Docker and Kubernetes runtimes behind one lifecycle interface: create, inspect, pause, resume, renew, snapshot and delete. A daemon called execd runs inside each sandbox to execute commands and code, handle files and report resource use. "While OpenSandbox was created at Alibaba, we don't want its future to be defined by a single company," said Xiaobin Xu, Alibaba's head of development infrastructure.

GitHub stars give a rough sense of how much developer attention each project brings with it:

What this means for your stack

• Write skills without a vendor bet. A SKILL.md folder is now a foundation-governed format with lead maintainers from the MCP side and contributors from Google and OpenAI. Pin to whatever the first versioned release turns out to be, and read skills before you install them.
• Code to one LLM API shape. If you already use the OpenAI Responses API, you are most of the way to Open Responses. Vercel AI Gateway and Databricks put several providers behind that one request format, and Claude and Gemini get there through translation layers for now.
• Treat sandboxes and environments as shared infrastructure. OpenSandbox started inside Alibaba and OpenEnv arrives through Hugging Face. Under AAIF both can pick up outside maintainers, which helps if you were wary of building on a single company's repo.

Key Takeaways

• AAIF added four projects on October 22, 2026, taking it from three at its December 2025 launch to 10.
• OpenAI donated Open Responses, an open LLM API spec based on its Responses API. Databricks and Vercel ship implementations, and Anthropic and Google are absent from its community list.
• Anthropic donated Agent Skills: more than 70 platforms, about 1 million skills published and nearly 280 million installs, maintained by engineers from several companies.
• Hugging Face's OpenEnv standardizes RL environments with reset(), step() and state(); Alibaba's OpenSandbox runs agent workloads on Docker or Kubernetes under Apache 2.0.
• The Agent Skills roadmap includes versioned releases and security guidance for skill installers, neither of which exists yet.

Sources: AAIF press release, Anthropic: Donating Agent Skills, AAIF: What Open Responses changes, AAIF: OpenEnv, AAIF: OpenSandbox, openresponses.org, Open Responses changelog, InfoQ, SiliconANGLE, Unite.AI, Simon Willison, Databricks docs, Vercel docs, agentskills on GitHub