A newsletter on coding agents tours the new software factories. One merges over a thousand agent pull requests a week. Another forbids humans from writing or reviewing code and budgets a thousand dollars a day in tokens per engineer. Asked how it was going, that one offered cadence and optimism but no results data. I read that line twice.
Pull requests per week is a speed number, not a quality number. The figure I want is the other one: how much of this code does someone still understand? That gap has a name: comprehension debt. The teams I trust put a human back where it is cheap: before the build, at the design of types and interfaces, and on slices of a hundred lines instead of two thousand at the end. So ask the factory for results. If all it shows is cadence, you are watching a demo.