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

TL;DR - This study models repository-scale coding as maintaining a graph of coupled facts, with unavailable facts creating “coherence debt.” Across seven models and five harnesses, fact availability—not context distance or extra token spending—primarily determined success.

• Missing facts caused agents to fabricate files or guess values rather than leave work incomplete.
• Prompt-supplied facts worked equally well whether near or far from the relevant edit.
• Successful harnesses differed by over 10× in token use because they reconstructed context at different rates.
• Stale conventions could be worse than absent guidance, while memorized repository knowledge weakened read-based success metrics.