Research·arXiv cs.LG·15h ago

Coding agents fail on repository tasks due to coherence debt, not context limits

A new study quantifies how coding agents fail at repository-scale tasks by modeling the problem as a graph of interdependent facts that must stay coherent across edits. Researchers tested seven models across five different harnesses, systematically removing access to recent context or learned parameters to isolate failure modes. The key finding: whether a required fact is available in the prompt or memorized determines success or failure, while its distance from the agent's current position is irrelevant. When models encounter unfamiliar APIs or library changes, they collapse uniformly at the same point. This work exposes a fundamental architectural constraint in agentic coding systems and suggests that scaling context windows alone won't solve coherence problems without addressing how agents retrieve and maintain coupled dependencies.

Modelwire context

Explainer

The study's core insight is architectural, not empirical: it shows that repository-scale failures stem from a retrieval problem (whether facts are accessible), not a distance problem. This inverts the common assumption that longer context windows solve agentic reasoning at scale.

This connects directly to the August gradient bottleneck work, which also challenged an assumed architectural constraint by isolating what actually limits performance. Both papers use causal intervention to separate forward-pass geometry from what practitioners blame. The coherence-debt framing also echoes the distillation paper's finding that origin relationships (here, whether a fact lives in prompt or parameters) matter more than surface-level metrics. Together, these suggest the field is converging on a pattern: blamed bottlenecks often misdirect attention from retrieval and dependency structure.

If teams building coding agents adopt coupled-fact-graph modeling as a standard design pattern within the next 12 months, that signals the paper moved from diagnosis to practice. Alternatively, if a major model release announces architectural changes to fact retrieval (rather than context scaling) in response to this work, that confirms the finding has shifted how practitioners think about agentic design.

Coverage we drew on

Does the LM Head Create a Harmful Gradient Bottleneck? A Causal Test · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

Mentionscoding agents · repository-scale tasks · coupled-fact graph · parametric memory

Read full story at arXiv cs.LG →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “The Working Set of a Coding Agent: Coherence Debt in Repository-Scale Tasks”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.