Measurement

Commits, lines, pull requests and tickets closed were always weak proxies, tolerated because producing them took human effort — the effort was the thing that made them mean anything. Agents remove that constraint, so the proxies now move whether or not anything improved, and they move most for the people delegating most indiscriminately.

A measurement that rises when you do the wrong thing faster is worse than no measurement, because it will be quoted.

For each candidate measurement, write down how you would inflate it if you were trying to. If you cannot think of a way, you have not thought hard enough. Then ask whether the inflating behaviour is one you would notice — and whether it looks different from real improvement.

The measurements that survive tend to share a shape: they count outcomes downstream of the work rather than the work itself, and they get worse when quality drops. Time from problem noticed to problem gone. Rework rate on accepted changes. How much of your week was repair. That third one you already have, from module 00.

Three is the limit on purpose. A dashboard of twelve is a way of never having to conclude anything.