Why does grading each agent separately miss multi-agent failures?

Because the failure often lives in what one agent’s message to another does or doesn’t carry, and each agent’s own output can look correct when graded alone. MAST’s own trace data shows this directly: a Phone Agent calls a login API with the credentials it was given, gets “invalid credentials” back, and reports that accurately to the Supervisor Agent, without saying the username field needs a phone number. Read either agent’s turn on its own and it looks reasonable: the Phone Agent described what actually happened, and the Supervisor Agent asked a sensible follow-up given what it knew. The login keeps failing anyway, because the one fact that would have fixed it never crossed the handoff.

That’s why inter-agent misalignment, 36.9% of the paper’s annotated failures, is defined at the boundary between agents rather than inside any single agent’s output. MAST classifies the trace, not the answer, because a grader scoring one agent’s final response on its own never sees the message that didn’t get passed.