Study Proposes Gating AI Coding Agent Refactors by Where Codebase Debt Costs Most

Coding agents make software faster to produce, but each change leaves a codebase that shapes the next agent run, so poor structure can spread. The article argues refactoring should be prioritized by recurring cost rather than by how messy code looks. In the authors' evaluations, reviewing every pull request for refactoring cost four to nine times as much per trajectory as doing no refactoring. Their approach uses signals such as file change frequency, cross-directory coupling, and static-analysis health to find hard-to-change code that is frequently touched. A deterministic gate runs before any model review, and a separate agent proposes refactors that require human approval and are checked by behavior-preserving replays.