AI agents are becoming API consumers in their own right. The good news: an API designed well for human developers is most of the way to being usable by an agent. The discipline you already know still applies.
Predictable beats clever
Agents struggle with surprises. Consistent naming, consistent error shapes, and consistent pagination matter more than ever. If every endpoint behaves the way the others do, an agent can generalize from a few examples instead of memorizing exceptions.
Describe yourself precisely
A complete, accurate OpenAPI spec is now doing double duty: it documents your API for people and it tells an agent exactly what each operation does and expects. Tight descriptions and explicit types reduce the guessing that leads to bad calls.
Make actions safe to retry
Agents retry, sometimes aggressively. Idempotency keys on writes turn a scary class of duplicate-action bugs into a non-issue. Clear, structured errors let the agent correct itself instead of looping.
You do not need a special “AI API.” You need a well-designed one — which you should have been building anyway.