Story · arXiv
The Working Set of a Coding Agent: Coherence Debt in Repository-Scale Tasks (arXiv)
paper · Story page
The paper models repository-scale coding as rebuilding a graph of coupled facts; facts covered by neither context nor parametric memory are coherence debt. Across seven models and five harnesses, availability decides the outcome, not distance from the edit, and passing harnesses differ more than tenfold in tokens.
In plain words
- A study found that coding systems fail when required project facts are unavailable, even when those facts concern nearby changes.
- Large software projects contain connected requirements involving tests, file references, settings, and rules for updating stored information.
- Before editing, a coding system must retrieve each relevant requirement from recent text or its learned knowledge.
- Supplying a missing fact restored success, while processing more text could not replace information that was never provided.
- Builders of coding tools should keep required facts available instead of assuming more text processing will solve the problem.
Appeared in
- Aborted agent branches live on in the KV cache, plus StateM's harness scaling
Aug 19, 2026 · in the sections
Subscribe
Get the brief in your inbox
Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.
- Weekdays at 8:45am IST, one lead story and 6 to 9 items.
- Sundays, an argued synthesis rather than a recap.
- One click to leave, and quiet days say so in the subject line.