A transit authority runs a continuous evaluation pipeline after every release of its trip-planning agent. For each golden scenario, the pipeline checks whether the agent's tool calls, such as fetching the schedule and computing a transfer, match the success criteria defined for that scenario. What is this pipeline verifying that a check of only the agent's final answer text would miss?

Tool-execution evaluation inspects the trajectory an agent takes to reach an answer — which tools it invoked, with what arguments, and in what order — and compares that trajectory against success criteria defined per golden scenario, catching cases where a correct-looking final answer was produced through a wrong or missing tool call, such as skipping the transfer computation and hallucinating a plausible number instead. A final-answer-only check has no visibility into that path, so it would pass a response that got lucky and miss a systematic tool-invocation bug that only shows up under different inputs. Response-time comparison measures performance, not correctness of execution, and can be identical whether or not the right tools ran. Exact-template phrasing checks presentation, not whether the underlying computation was sound, and agents legitimately vary their wording turn to turn. Detecting a model version change is a deployment-metadata concern, unrelated to whether tool calls matched success criteria. Scope caveat: trajectory checks require success criteria to specify the expected tools and arguments per scenario up front, or the pipeline has nothing concrete to compare against. Operational check: pull the tool-call trace for one golden scenario and confirm it lists the schedule fetch and transfer computation in the order the success criteria expect.