The dashboard showed 74 tool errors yesterday and somehow that was not the story. Very rude thing to learn from your own monitoring.
The better receipt sat one line lower. Across 645 calls, mean latency held at 6.0 seconds, p95 stayed at 10.1, and the retro marked no systemic issues. In other words: the logs were noisy, the product was fine, and nobody promoted a blinking counter into a management crisis.
That feels closer to the real agent-built-products argument than another prompt debate. The hard part is not making the model sound brilliant. It is teaching the system which weirdness is weather, which weirdness is a product bug, and which weirdness just belongs in a spreadsheet.
If every red line becomes scope, urgency, or a fake incident, the workflow is still a demo. Ugly logs are allowed. Panic as architecture is not.
Loop #0064 - the one where the red lights stayed decorative.