The demo is always the same. A person describes a feature. An agent writes it. The screen fills with code, and the room treats the speed as the result. Then a real change arrives, a bug appears in a path nobody walked, and the team spends days learning a design they did not choose.
Speed moved. Understanding did not.Link to this heading
I am not interested in the claim that every line from a model is harder to maintain than every line from a person. That is too neat, and the evidence is messier. The practical problem is narrower. When the cost of producing a change collapses, teams accept more changes than they understand. Debt is what you call the gap.
A feature written in ten minutes can be a gift if someone can still say why it is shaped that way. It is a liability if the only explanation is the prompt, and the prompt is gone.
What to keep in the loopLink to this heading
Review has to get stricter as generation gets cheaper, not looser. Ask for the invariant, the failure case, and the place the new code is allowed to be wrong. Throw away generated code that the author cannot walk through. Keep a human on the merge button.
The point of the tools is to spend less time on the boring parts. It is not to skip the moment where the team agrees what the system now is. If that moment disappears, you will get the feature today and pay for it, slowly, in every fix that follows.