REST made software legible to developers by turning capabilities into predictable resources. Agents need another layer. They need to understand what a capability means, when it is appropriate, what it will change, and how to recover when the world does not match the expected state.

That semantic contract cannot live only in prompt prose. It belongs close to the system of record: typed intent, policy, identity, provenance, observable side effects, and explicit recovery. The best agent interfaces will feel less like a bag of API calls and more like a trustworthy operating model for software.

The engineering opportunity is not to replace APIs. It is to make the meaning around them executable—so agents can act with the same context a strong human operator would demand.