Integration Testing for Flows

Unit tests prove each tool works alone; integration testing proves the whole agent flow works together. Here you run a realistic task end to end — user message in, model reasoning, tool calls, final answer out — and assert on the outcome and the trajectory. Instead of checking one function’s return value, you’re checking that the pieces compose: the router picks the right sub-agent, tool outputs feed correctly into the next prompt, and the conversation ends in a valid state.

This matters because most agent bugs live between components. Every tool can pass its unit tests while the flow still loops forever, calls tools in the wrong order, or loses context mid-conversation. LLM systems are also non-deterministic, so a flow that worked yesterday can regress after a model update or prompt edit with no code change at all. Integration tests are your safety net for exactly those failures — the ones users actually see.

In practice, you’ll build a suite of scenario tests: a fixed set of realistic tasks with clear success criteria. For each, assert on final outcome (did the refund get issued, does the answer contain the right facts) and on trajectory (which tools were called, in what order, within a step budget). Mock external side effects so runs are repeatable, allow multiple valid paths, and score with a mix of hard assertions and LLM-as-judge grading. Run the suite in CI against recorded fixtures, and track pass rate over time rather than demanding 100% on every run.