Tech giants unite to create open standards for AI agents. OpenAI donates AGENTS.md, Anthropic contributes Model Context Protocol, and Block adds goose framew...
What the Foundation Actually Standardizes
The new agentic AI foundation brings OpenAI, Anthropic, Google, Microsoft, and Block together around a shared goal: open standards for how AI agents are described, connected, and run. Rather than each company shipping its own incompatible way of wiring a model into tools and files, the foundation pools existing, already-proven pieces into one governed home. That matters because an agent is only as useful as the systems it can reach, and reaching those systems has so far meant custom glue for every model and every vendor.
The initial contributions map cleanly to the two hard problems in agent building. OpenAI donates AGENTS.md, a convention for telling an agent how a given project or codebase expects to be worked with. Anthropic contributes the Model Context Protocol, which defines how an agent connects to external data sources and tools. Block adds the goose framework, a runtime for actually executing agent workflows. Together they cover instruction, connection, and execution.
Why Shared Standards Beat Private Ones
When every provider invents its own agent interface, integration cost multiplies. A team that builds a tool connector for one model has to rebuild it for the next, and switching providers means rewriting the plumbing rather than swapping an endpoint. Open, commonly governed standards break that lock-in: a connector written once can serve any agent that speaks the same protocol, and a project's AGENTS.md guidance applies no matter which vendor's agent reads it.
Putting competing companies behind a single foundation also changes the incentives. A standard maintained by one firm can be steered to favor that firm's products; a standard held in common has to stay neutral to keep its contributors invested. For developers, that neutrality is the whole point — it is what lets them commit to a format without betting on which company wins.
What This Means for Builders
If you are building or integrating agents, the practical guidance is to align with these standards early rather than growing a private convention you will later have to unwind. A few concrete steps follow from the foundation's structure:
- Add an AGENTS.md to your repositories so any compliant agent inherits your build, test, and style expectations without bespoke prompting.
- Expose your internal tools and data through the Model Context Protocol instead of one-off, model-specific adapters.
- Evaluate a runtime like goose for orchestration so your execution layer is portable across the models you might adopt.
The payoff is that each layer becomes replaceable. You can change the underlying model, the tools it calls, or the runtime that drives it without tearing up the other two, because the seams between them are defined by shared specifications rather than by any single product.
The Open Governance Question
Bringing direct competitors under one roof raises an obvious tension: whose priorities shape the standard when their commercial interests diverge? The strength of a foundation model is that no contributor can quietly break compatibility to advantage itself, but that only holds if governance stays genuinely open and the specifications remain implementable by anyone, not just the founding members.
The signal worth watching is adoption outside the founders. Standards earn their status when independent tools, smaller vendors, and open-source projects build on them freely. If the donated pieces stay open and widely implemented, the foundation becomes shared infrastructure; if they drift toward the interests of a few, it becomes branding. The contributions so far point the right way.