Short answer: Give the agent a local mock generated from the spec, not a live backend. It is deterministic, contract-shaped, and returns the edge cases on demand, with no account and no risk to shared data. The agent develops against the same contract the tests assert, and the mock regenerates when the spec changes.

Why a live backend is the wrong target

An agent hitting a real server mutates shared state, hits rate limits, and is slow. Worse, it cannot reproduce the rare failure: the 429, the empty list, the malformed response. The agent optimizes against the happy path because that is all the server reliably offers.

What a generated mock provides

Determinism: the same request returns the same response, so the agent can iterate.

Edge cases: error and empty responses come from the contract, not luck.

Safety: no shared data is touched and no network is required.

Sync: the mock is generated from the spec and cannot drift from it.

The workflow

Start the mock from the studio, point the agent at the…