Motivation

I have always liked types and tests for one reason: they let me stop thinking.

You think hard once, before writing the code: what should this endpoint do, what should it return, and to whom. You write that down as a test. From then on the test holds the behavior, and writing the code is almost mechanical. If something breaks, the test and the compiler point at the exact place, and you fix it without keeping the whole system in your head.

API tests rarely work like that. They talk to the backend through bare fetch: tokens passed by hand, ids kept in variables, any in every response. A renamed field shows up as a failed run, if it shows up at all. eberly brings back the “think once” part: the client comes from the spec, so the compiler knows every endpoint, status and body, and every test user keeps their own tokens and ids, so a test is just the scenario.

Then AI agents started writing code, and it turned out they have the same problem people do. Overload the context and an agent gets worse at the task, just like a person who has to keep too much in mind. The approach that takes load off a person’s attention fits an agent almost exactly:

• The agent doesn’t need the backend repository. The URL of a running backend and its OpenAPI spec are enough.
• It doesn’t load the whole spec either. It reads the endpoint index and opens only the endpoints of the test it writes.
• When it gets something wrong, TypeScript says where, and it fixes that one spot.

So eberly was useful before agents and is useful now for the same reason: attention is limited, whether it belongs to a person or a model.