my ai report generator agent demos like magic. 90% of the actual code is there because it lies confidently

I build agents for a living and the gap between the demo and the thing you can leave running still surprises people, so here's the honest breakdown. I built an agent that pulls data from a few systems and writes a summary report. In the demo it looks like magic. You ask, it thinks, out comes a clean report. Everyone's impressed. That part is maybe a tenth of the code. The other ninety percent is there for one reason: the model fails silently and confidently. It will invent a number that looks exactly as plausible as a real one. It will summarize a table it half-read. It will call a tool, hit a rate limit, and cheerfully write the report as if the data came back. So most of what I actually wrote isn't "the agent." It's: Retries with backoff for every tool call, because half the failures are transient and the model has no idea. Output checks that reject the response if a required field is missing or a number doesn't reconcile against the source, before a human ever sees it. A hard rule that if a data source didn't return, the agent says "I couldn't get X" instead of guessing. Getting it to admit the gap instead of papering over it was most of the work. Logging every decision so when it does go wrong i can trace which step lied. The demo sells the 10%. The 90% is what decides whether a client trusts it in month two or quietly turns it off. And none of the 90% is impressive to watch, which is exactly why the flashy threads never show it. For people running agents in production: where do yours fail silently, and what's the check that finally caught it? Feels like everyone rediscovers output validation the hard way.