Conceptual integrity and counting lines of code

Simon Willison has been making a fresh case for a claim many engineers still hate: sometimes lines of code do tell you something. On a recent Talking Postgres episode with Claire Giordano, he argued that coding agents can push a senior engineer from a few hundred lines of production-ready code a day to something far beyond that. Not sloppy code, either. The point is that if the output is debugged, tested, and maintainable, the jump is real.

He’s careful about where the argument stops. A person can produce much more code with agents, but not 100 times more cognitive capacity. That becomes the new bottleneck. One engineer can now churn out far more software than before, yet still can’t keep that much of it in their head at once. So the team doesn’t disappear; it just shifts from pure code production to sharing the mental load.

Willison’s other warning comes from a very different place: The Mythical Man-Month and its idea of conceptual integrity. Good software should feel coherent, with parts that fit and nothing odd hanging off the side. Agents make that harder. You can prompt for a feature, wait five minutes, and suddenly the codebase has another little bump in a new direction.

Claire Giordano’s Winchester Mystery House comparison fits because it captures the danger neatly. If adding a room is cheap, you keep adding rooms. The real discipline used to come from time and pain; a wild feature that would take a week was easy to reject. When it takes an hour, the same bad idea gets much harder to kill.

That’s the sharp part of Willison’s argument. Cheap code is not the same as cheap thinking, and AI tools may be exposing that gap in the ugliest possible way.