Agents removed the "I got lost in this code" signal that used to trigger refactoring.

A September 2 essay argues the human refactoring reflex was fired by cognitive strain, and since agents navigate tangled codebases without that strain, nothing flags accumulating complexity until no developer on the team can reason about key parts and reviews degrade into rubber stamps. The author is explicit that this is anecdotal with no measurements behind it. The economic turn is the interesting bit: clean modular code cuts token cost and hallucination rate, so refactoring pays off for the agent even where no human reads the code. I've felt exactly this and I don't have data either.