The cost of a half-finished abstraction
Unfinished abstractions are worse than none, and agents make them cheaper to create and harder to spot.
A missing abstraction is annoying in a way that is easy to see: the same logic sits in four places and you have to remember to change all four. A half-finished abstraction is worse, and it is much harder to see.
The shape is always similar. Someone extracts a common piece, covers three of the five call sites, and leaves the other two doing it the old way. Now the codebase has two truths. A reader finds the abstraction, assumes it is authoritative, and is wrong. The next person extends the abstraction for their case, which makes it slightly more general and slightly less honest about what it actually handles.
The cost is not the duplication. It is that the abstraction makes a promise the code does not keep, and every subsequent reader pays to discover that.
What changed recently is the economics. Extracting an abstraction used to be slow enough that you thought about it first and finished it once you started. With an agent it is fast, which means half-finished ones get created more easily. They are also harder to spot in review, because the diff shows a clean extraction and says nothing about the call sites that were left behind.
The only countermeasure I have found so far is mechanical: when an abstraction is introduced, ask for the list of call sites that were not migrated, in the same change. If that list is not empty, the change is not done. I do not love this, because it relies on me remembering to ask.