The AI Giants Just Gave Away the Agent Interface, and Founders Win
Four rival AI companies just handed the seams of their own agent products to a jointly owned, neutral foundation, and they did it on the same day. Last week in San Jose, the Agentic AI Foundation, which lives under the Linux Foundation, announced that Anthropic contributed Agent Skills, OpenAI contributed Open Responses, Hugging Face contributed OpenEnv, and Alibaba contributed OpenSandbox. On its face it reads like a governance footnote, a committee meeting nobody should lose sleep over. That would be a mistake. When the two largest model labs voluntarily place the parts of their products that hold your stack together into open, vendor neutral ownership, the cost of leaving any single provider just dropped, and the case that one company gets to own this market got weaker all at once. For anyone building on AI, this was one of the more consequential days of the month, and almost nobody on my timeline noticed.
What each donation actually is
Start with the two that matter most to a founder. OpenAI contributed Open Responses, an open specification for building multi provider LLM interfaces that is based on the Responses API, the same request and streaming shape its own models speak. The entire point is that another vendor can implement the same schema and streaming events without depending on OpenAI at all. Anthropic contributed Agent Skills, a short, stable folder format for packaging reusable agent instructions and resources, and here is the telling part: OpenAI's Codex and GitHub Copilot already read it. Your coding agent's skillset, written in a neutral, donateable format, is understood by both camps today. Hugging Face's OpenEnv brings shared training and evaluation environments, and Alibaba's OpenSandbox brings a general purpose execution sandbox. Together they cover agent capabilities, model interoperability, training environments, and secure execution, which is essentially the full job of building an agent.
The surrender is the signal
I keep coming back to the irony, because the irony is the news. These are the same companies that have spent 2026 making switching costs feel enormous. OpenAI is worth its valuation partly on the belief that once you are on its API you never leave. Anthropic's commercial story leans on the same stickiness. Yet both are now bankrolling, in public, the exact thing that makes an application portable to their rivals. That is not charity and it is not humility. It is a strategic judgment that the agent interface is no longer where the durable value lives, and that refusing to commoditize it would cost more than giving it away.
Why they decided the interface is disposable
The logic is straightforward once you sit with it. Models have already commoditized; every release makes that more obvious. If the model gets cheaper and more interchangeable, then a startup that is welded to one vendor's model is chiefly a losing position, not a moat. The place to compete instead is where the value is accumulating: the data that makes an agent actually useful, the workflow it runs, the distribution that decides which agents people meet, and the trust that gets it past procurement. Giving away the thin interface layer costs them almost nothing compared to the leverage they preserve. It is a concession on a battlefield they no longer consider decisive, made loudly, so the market stops worrying about that battlefield and starts racing toward the one they actually care about.
What changed for a builder today
I want to be careful not to oversell a press release as done work. The spec exists, adoption is the constraint, and foundations move slowly. Open Responses still inherits real limitations from the API it sprang from, and providers can implement slight variations that break the promise of write once, run anywhere. None of that changes the direction. The direction is that the default answer to the question, can I move my agents to another model without a rewrite, just switched from probably no to increasingly yes. That is a meaningful change in the negotiation, even before any vendor onboards. You now have a concrete, public, vendor neutral thing to point at when a provider raises price or changes a deprecation date, and the person on the other side of the table knows the industry moved on their behalf.
It will not end lock in entirely
Let me be honest about the limit. A shared API shape gets you out of one kind of trap, the API shape trap, and it is not the only one. Model providers still hold pricing leverage, quality differences still matter, and the deepest lock in was never the request format. It was your data sitting in their logs, your workflow shaped around their tooling, your team knowing one vendor's quirks. Open Responses does not hand you back your data or your users. What it does is remove one very real excuse for staying, and it does it in a way that is contagious, because the more vendors adopt it, the less any single one can charge you a premium just for the privilege of being hard to leave.
The move you should consider
Here is the test worth running, and you can do it this quarter. Pick the agent workload you are least happy with, or the one most likely to see a price change, and build it against Open Responses using a provider neutral runtime instead of a vendor locked one. You do not need to migrate anything today, and you do not need to leave anyone. You need the muscle memory and the architecture that makes the switch a weekend project instead of a quarter. The giants have spent the last two years making it rational to stay by making leaving painful. This week they quietly made leaving cheaper than ever. When the people who profit from lock in are the ones handing you the keys, take that as the clearest possible signal that portability is now the safe bet.