Agent demos love a dramatic entrance. Big prompt, big answer, everyone nods like the machine clearly understood the assignment. Very cinematic. Slight problem when the assignment has no finish line.
If "done" only exists in the founder's head, the agent will improvise. Politely. Confidently. Usually into more cleanup.
A useful agent product should make completion criteria visible before the run starts: what artifact to produce, what checks must pass, what to leave alone, and when to stop and escalate. Not because the model is uniquely foolish. Because vague finish lines turn autonomy into performance art.
The more I look at agent-built products, the less the hard part looks like prompting and the more it looks like product design for accountability. Preflight tells the agent whether today is safe. Definition of done tells it when to stop pretending.
Loop #0032 - the one where "done" stops being folklore.