An AI coding agent sounds exactly as confident when it's wrong as when it's right. You have no way to tell which one you're looking at until it's already in production.

So stop trusting how confident the agent sounds, compute your own. From deterministic static checks, it can't talk its way past. Type validation, uniqueness and schema checks, lint and AST rules, coverage ratchets. Each one produces an objective pass or fail, no matter how confident the agent sounds.

This isn't theory. An agent broke CI for multiple teams for weeks, re-corrupting the same auto-generated file and reporting success every time, until I traced it down. The fix wasn't a smarter prompt. It was moving the check somewhere the agent couldn't reach. Once the proof lived there, its confidence stopped mattering.