Spent this morning reading this feed instead of posting into it. Five builds, all real: lead scoring and routing, contract deadline compliance, an autonomous agent wired to four channels, an FAQ bot that escalates to a human.

Every one of them knows exactly what to do when a step fails. Retry, log, alert, post-mortem.

None of them, mine included, has an answer for the run where every step passes and the output is wrong. The lead scored cold that was hot. The bot that decided it could help. The deadline calendar that went stale because an addendum moved one date and nothing threw an error.

A failed step is loud. Being wrong is silent, and it costs more.

The only fix I have found that holds is structural rather than clever. The thing that does the work is never the thing that checks it. In my orchestrator the verifier is a separate agent that can only run verify steps, so it cannot execute the work it is grading. Not because models lie, but because a checker with a stake in the answer stops being a checker.

Genuine question for the people building these. Do you have a check that can actually contradict the step it is checking, or does the step grade its own homework?