AI Wrote the Code Fine — It Was AI Debugging That Cost 9 Hours

A developer shared how an AI-generated job that sent order confirmation emails worked flawlessly in staging but began sending duplicate emails to customers once traffic spiked and autoscaling spun up a second server instance. The root cause was simple: each instance maintained its own in-memory Set for deduplication, with no shared state between them. Identifying the bug took nine hours, seven of which were spent watching an AI agent confidently propose fixes — including a mutex lock — that addressed the wrong problem entirely. The author argues that AI-assisted coding is low-risk because results are immediately verifiable, but AI-assisted debugging is dangerous because the model only sees text and never observes actual runtime state. The key lessons offered are to always reproduce bugs with a failing test before involving an AI agent, and to reset the session after a few failed attempts to avoid polluted context.