The Agentic AI Foundation (AAIF) picked up four new projects on Thursday, October 22, and the donor list reads like a frontier-lab group photo. Per AAIF’s announcement, made at AGNTCon + MCPCon North America in San Jose, the new projects are:
- Agent Skills, from Anthropic: an open format that packages instructions, scripts and resources that agents load on demand.
- Open Responses, from OpenAI: an open spec for talking to language models from many providers, based on OpenAI’s own Responses API.
- OpenEnv, from Hugging Face: a framework and hub for building reinforcement-learning (RL) environments, the simulated tasks used to train agents by trial and reward.
- OpenSandbox, from Alibaba: a sandbox platform with SDKs and Docker and Kubernetes runtimes for coding agents, GUI agents, evaluation and RL training.
That takes AAIF from three projects at its December 2025 launch to 10. The others are the Model Context Protocol (MCP), AGENTS.md, goose, the Agent2Agent (A2A) protocol, agentgateway and Agent Router.
What AAIF is, briefly
AAIF sits under the Linux Foundation, which has become the default neutral home for big shared open-source projects. The New Stack notes the foundation launched with three founding projects: Anthropic’s MCP, the standard for connecting agents to tools and data; OpenAI’s AGENTS.md, a file that gives coding agents project-specific instructions; and goose, an open-source agent from Block.
Donating a project to a foundation means handing over the trademark and governance. In practice, technical decisions move to a steering committee, which The New Stack says will initially be made up of project maintainers. The company that wrote the code no longer gets the final say, at least on paper.
Why Open Responses is the interesting one
Open Responses defines “a shared schema, consistent streaming events and extensible tooling,” so developers can describe a request once and run it across model providers “with minimal translation,” AAIF says.
OpenAI first announced the spec in January. The idea, per The New Stack, is that every provider now offers the same building blocks (messages, tool calls, streamed output) but represents them differently. Switching models means rewriting the glue.
Support is already broad. The New Stack lists AWS Bedrock, Databricks, LM Studio, Ollama, Hugging Face, Nvidia, OpenRouter, Red Hat, Vercel and the vLLM project among implementers. AWS says it employs a core maintainer on the project and that Bedrock has supported the Responses API format since December 2025.
Our read: the most useful thing about this is also the most self-serving. If the industry’s common interface for talking to models is a cleaned-up version of OpenAI’s API, then OpenAI’s API is the default and everyone else is the adapter. The New Stack points out that the project’s own contributor guide says its core schema files are copied from OpenAI’s API. That’s not a scandal so much as a description of how standards usually win: being first and being donated beats being better.
The catch: two major names are missing. Anthropic still uses its own Messages API, and Google has its Interactions API, according to The New Stack. Anthropic’s format has momentum of its own: DeepSeek, Moonshot, Z.ai and MiniMax expose Anthropic-compatible endpoints, largely thanks to Claude Code’s popularity. So the “standard” for calling models is currently at least two standards.
Anthropic’s parallel move
Anthropic is playing the same game with Agent Skills. A skill is essentially a folder: instructions, scripts and reference files that an agent loads only when a task needs them, rather than stuffing everything into every prompt.
“We launched Agent Skills a year ago to give developers and enterprises alike an open standard to make their agents more capable,” said David Soria Parra, co-creator and lead maintainer of MCP, in AAIF’s release. He pointed to AAIF’s stewardship of MCP since Anthropic donated it “less than a year ago.”
So each of the two biggest US labs has now handed the foundation one piece of connective tissue while keeping its own model API. OpenAI gave away the request format; Anthropic gave away the packaging format for know-how.
Why it matters: labs compete hard on models, then give away the plumbing (how agents get tools, skills, environments and a safe place to run) so customers don’t have to rewrite integrations per vendor. Lower switching costs sound bad for incumbents. But they also make enterprises more comfortable building on agents at all, which grows the market every lab is selling into.
The training and sandbox pieces
The two less-noticed projects may matter more to builders.
OpenEnv packages RL environments as containerized services that speak both MCP, the interface agents already use, and Gym, the long-standing interface RL training libraries expect. That means an environment built for evaluating an agent can, in principle, also be used to train one. Hugging Face’s Ben Burtenshaw listed Hugging Face, Microsoft, Reflection, Meta-PyTorch, Fleet, Mercor, RadixArk, Modal and Prime Intellect among groups that built and used it.
OpenSandbox addresses where agents run. Coding agents execute code they wrote themselves, and you don’t want that happening on a production server. OpenSandbox defines lifecycle and execution APIs so teams can plug in their own runtimes. “While OpenSandbox was created at Alibaba, we don’t want its future to be defined by a single company,” said Alibaba’s Xiaobin Xu.
The catch: a Chinese tech giant’s sandbox joining a US-heavy foundation is a nice signal for openness, but enterprises with export-control or supply-chain policies will still review it on their own terms. Neutral governance doesn’t change who wrote the first version of the code.
Who this affects
- Developers building agents: fewer bespoke integrations, if the specs get real implementations. Writing against Open Responses today already works on Bedrock, Ollama and OpenRouter, per the implementer lists above.
- Model providers: pressure to support a shared format, especially smaller providers that benefit most from easy switching.
- Anthropic and Google: the obvious holdouts on the request format. Anthropic is an AAIF member, as The New Stack notes, so adoption isn’t out of the question.
- Enterprises: procurement and governance teams like standards run by foundations rather than single vendors. This gives them more of those.
The quotes, for the record
“Four different companies chose to contribute them to neutral governance,” said AAIF CTO and co-founder Manik Surtani, adding that the foundation’s job is giving the community one place to build standards “with no single company controlling the outcome.”
OpenAI’s Steve Coffey said Open Responses is joining to help “developers build great products on today’s language models through simple, clean, and unified interfaces.”
The takeaway
The takeaway: the agent stack is consolidating around a handful of open specs, and a growing share of them now live under one neutral roof. That’s genuinely useful. It’s also strategically convenient for the companies whose formats became the specs.
What to watch:
- Whether Anthropic or Google ship Open Responses-compatible endpoints, or whether the Anthropic-style format keeps spreading in parallel.
- Whether Open Responses gets compliance tests that providers actually pass, rather than “mostly compatible” claims.
- Whether OpenAI’s own newer endpoints, like its Decisions API, end up in the spec or stay proprietary.
- Real adoption of OpenEnv and OpenSandbox beyond their founding contributors.
Image: Michael Ocampo / Wikimedia Commons, CC BY 2.0
