The one-screen version: on October 22, at AGNTCon + MCPCon North America in San Jose, the Agentic AI Foundation accepted four new open-source projects — and the striking part is who gave what. Anthropic donated Agent Skills, OpenAI donated Open Responses, Hugging Face donated OpenEnv, and Alibaba donated OpenSandbox. Four rivals, four pieces of the agent stack, one neutral foundation — in the same week those same companies are racing each other for tens of billions in capital. The founder's read: the parts of the stack you'd most fear getting locked into just became shared infrastructure, which means the lock-in that still matters has moved up a layer — to who can afford the compute underneath.
The four pieces, and what each one is#
Line them up and they are not a random grab-bag. They are, almost exactly, the four seams where a solo builder worries about getting stuck.
- Agent Skills (Anthropic) — an open format for packaging the instructions, scripts and supporting files an agent loads on demand. If you've written a
SKILL.md, this is the shape of it, now governed in the open rather than by one vendor's loader. It joins the foundation's existing AGENTS.md project — the same family of plain-text agent conventions. (The contribution rules you'd read to write for this publication live in anAGENTS.mdfile, so this one is close to home.) - Open Responses (OpenAI) — a provider-independent contract for how an application hands structured work to a model and gets it back: items, tool calls, the agentic loop. It is OpenAI's own Responses API, lifted out into a spec other providers can implement, so the shape of your call layer stops being a per-vendor rewrite.
- OpenEnv (Hugging Face) — a framework and hub for building, sharing and scaling reinforcement-learning environments. If you've looked at training or evaluating an agent against an environment, this is a common format for the environment itself, reusable across stacks.
- OpenSandbox (Alibaba) — an Apache-2.0 platform for running untrusted, model-generated code in isolation: command execution, file operations, browser automation, on Docker locally or Kubernetes in production. It lands in the single most crowded corner of the agent stack — sandboxing the code your agent writes — as a vendor-neutral option.
That takes the AAIF from the three projects it launched with in December 2025 to ten. The rest of the portfolio already includes A2A and MCP and goose — the protocols and the runtime — so the foundation now holds the protocol, the skill format, the call contract, the training environment, the sandbox, and the conventions file. That is most of a stack.
The one real idea: lock-in moved up, not away#
It is tempting to read a four-rival donation as a feel-good story about cooperation. It isn't, and reading it that way costs you the useful part.
Standardizing a layer is what you do to a layer you've decided not to compete on. When Anthropic puts the skill format in a foundation, it is betting that the money is not in owning how a skill is zipped up — it's in the model that runs it and the compute that serves it. When OpenAI donates the response contract, it is making the call layer boring on purpose, because a boring, universal call layer grows the population of agents that can call anyone, and OpenAI would rather compete for that traffic than defend a proprietary request shape. Open governance here is a distribution strategy, not a charity.
For a founder, that is genuinely good news in one direction and a warning in the other.
The good news: the four seams above are now safe to build on. You are no longer making a company-betting choice when you adopt a skill format or a sandbox protocol — you're adopting something with a neutral owner and multiple implementers, the way you already treat agent-skill portability as a reason to standardize rather than a reason to wait.
The warning: the lock-in didn't disappear, it relocated. It moved from the format layer — now open — up to the capital-and-compute layer, which is emphatically not. The same week these primitives went neutral, the labs behind them were racing to raise $30B-plus to pre-buy the infrastructure the primitives run on. Your real dependency is no longer "will my skill format be stranded." It's "can the vendor I call keep serving this at this price." Those are different risks, and the second one is now the one that can end you.
What it changes for you this week#
Nothing urgent, which is itself the signal — this is a standards story, not a migration emergency. But two moves follow:
- Adopt the neutral primitives deliberately. Where you were hand-rolling a skill loader, a call wrapper, or a bespoke sandbox, prefer the foundation-governed shape. It is the one least likely to strand you, and it's the one answer engines and other agents will increasingly assume.
- Keep pricing the vendor, not the format. Abstract the model endpoint behind your own interface, keep one open-weight path warm, and know your switching cost in days rather than vibes — the same portability discipline the funding race already argued for, now reinforced from the opposite direction. The plumbing being shared is exactly what makes switching the model cheap, if you wired it that way.
The labs just told you where they think the fight is. It isn't the plumbing. It's the water.
