Parallel LLM agents beat linear search for multi-file code fixes

A new arXiv paper from researchers Akeela Darryl Fattha, Kia Ying Chua, Lingxiao Jiang, and Laura Wynter challenges the default linear exploration structure used by most LLM-based agents when localizing files for multi-file code changes. The team argues that visiting one directory or file per step is a structural mismatch for issues spanning multiple subsystems. They propose a non-linear, domain-scoped parallel agentic exploration system, where multiple specialized agents are spawned to explore different code domains simultaneously.

Using SWE-bench Pro as an initial benchmark and focusing on the Ansible project, the researchers built a persistent-session evaluation framework for GitHub issues anchored at a single base commit. Their domain-scoped parallel agents—powered by a small Haiku-class model—achieved the highest micro F1 among Haiku-class models by a large margin on their expanded benchmark incorporating PRs from 2025 and 2026. Only the much larger Codex 5.5 High outperformed it. On the original curated 2020 SWE-bench Pro, a larger Sonnet plain LLM baseline scored higher precision by predicting fewer files, but at significantly lower gold recall. The paper also reveals three important findings: documentation evolution remains an unresolved latent dependency; naive file system access can actually degrade localization due to test-file over-prediction; and forced multi-agent consultation does not measurably help and substantially increases token costs. These insights suggest that smarter exploration structure, not just bigger models, can dramatically improve AI-driven software debugging efficiency.