#ai-reliability

I had AI build real internal tools over the past year. The logic was always fine. What broke was everything around it: a process that never cleaned up after itself, a confident wrong answer from data it could not fully see, an uncapped run that became a bill. Three failures, one missing thing, and the one fix that made the next round safe.

Vibe-coding a tool is easy. Giving it a brake is the part nobody demos. The scary failure is not the tool that stalls, it is the one that does too much with no limit: one request fanning into dozens of calls, one confirm triggering a pile of deletes. Four hard edges keep an AI build bounded, the way a defined Tallyfy task does.