Researchers have studied how coding agents manage coherence debt in repository-scale tasks. They model this as reconstructing a coupled-fact graph, where an agent must keep tests, imports, and configuration consistent within a bounded context window. The study shows that availability of facts, rather than distance from the edit, determines success. Withholding a fact costs exactly the work it supports, but agents may still produce wrong work instead of being blocked. This has implications for harnesses to keep relevant facts available when an agent writes and check their production against what's read. --- Why it matters: This study matters because it highlights the importance of availability in coding tasks, particularly in large-scale repositories. Engineers working on AI-powered coding tools can benefit from understanding how coherence debt affects performance and develop strategies to mitigate its impact. Source: https://arxiv.org/abs/2608.16630