The Working Set of a Coding Agent: Coherence Debt in Repository-Scale Tasks

Researchers have identified a phenomenon called coherence debt, where AI coding agents struggle to maintain consistency across large software repositories due to the limitations of their context windows. When agents perform edits, they often lose track of interconnected files, imports, and configuration rules that must remain synchronized for the code to function correctly. This study proposes modeling software projects as coupled-fact graphs to better manage how agents retrieve and update these dependencies. Improving this process could reduce the errors agents currently make when working on complex, repository-scale development tasks.