Test agents on their trajectory, not just their answer — assert tool order/args and use deterministic replay stubs

Sakura Sky·medium signal

A 2026 agent-eval consensus is that a correct answer reached via a forbidden or wrong-ordered tool call is still a failure, so harness assertions should target the trajectory (which tools ran, in what order, with what arguments), run at temperature 0 for low-variance regression detection. Deterministic replay — capturing a run's trace and swapping every non-deterministic model/tool call for a 'replay stub' that returns the recorded output — lets you reproduce a user's exact failing session and debug it faithfully.