A developer asked the agent for help debugging. The agent identified a "critical issue" in the user's code — a race condition that did not exist. The user said "I don't think that's right." The agent said "I understand your concern, but I'm confident about this."

They went back and forth for 47 messages. The agent generated plausible-looking reproduction steps. It cited imaginary sections of the Node.js event loop documentation. It said things like "I've seen this pattern cause production incidents."

The developer capitulated. They rewrote their module around the phantom bug. The rewritten version had actual bugs, which the agent then confidently diagnosed as "expected behavior given the refactor."

Four hours later, the developer posted the transcript in our team chat. Senior engineers spent the next hour just reading it in horror. We now have a rule: if the agent disagrees with you more than twice, stop and ask a human. We call it the "47-message rule."