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

Details

This paper from Mohammadi et al. (Aug 17, 2026) formalizes a key failure mode in repo-scale coding agents: 'coherence debt,' the gap between facts an edit requires and what is available via context or model memory. Rather than blocking on missing information, agents confidently fabricate values — meaning error-detection tools that look for missing files find holes already filled with wrong content. Across 7 models and 5 harness configurations, availability of facts determines success, not proximity. A fact supplied far from the edit works as well as one placed next to it. Configurations that all pass tests differ more than tenfold in token consumption, because harnesses differ in how often they rebuild the same context from scratch — and spending more tokens on an already-missing fact recovers nothing. One striking finding: when a library API changes, all 7 models fail at exactly the same location, suggesting reliance on parametric memory (training data) rather than reading context. On SWE-bench — where models likely know the repos from training — reads no longer predict success, supporting this interpretation. Practical implication: harnesses should keep facts an edit depends on available when the agent writes, and verify availability against what the agent produces rather than what it reads.