How do I map a production trace to a MAST failure mode?
Match the trace to a stage first, then to a mode inside it. A design flaw shows up before any real interaction: the same step running twice with nothing new learned between attempts, or the run continuing after the task is already done or already correctly ruled unsolvable. A breakdown between agents shows up mid-run: one agent’s own reasoning reaches a conclusion its message to the next agent doesn’t carry, or it acts on an ambiguous instruction instead of asking. A verification miss shows up at the end: a fix reported as applied with no check that it actually took effect, or a check that ran but only confirmed something trivial.
Don’t expect one mode to explain a trace on its own. The paper’s own correlation analysis finds specific failure modes correlate as high as 0.63 when they share a surface symptom, meaning the same symptom, information one agent needed and didn’t get, can register as a design flaw, a withheld message, or an ignored one depending on which agent actually had it and what it did next, the same tangle behind an error one agent makes becoming the next agent’s ground truth.