AutoCodeRover extends multi-agent collaboration into real-world code repositories. It orchestrates specialized agents that autonomously navigate a codebase, edit source code, and validate the changes, across complex projects spanning multiple files.
The paper that introduces it, AutoCodeRover: Autonomous Program Improvement, presents it as an automated approach for solving GitHub issues to autonomously achieve program improvement — the software maintenance and evolution work its authors distinguish from coding itself. In that account, LLMs are combined with sophisticated code search capabilities, ultimately leading to a program modification or patch.
Capabilities
The introducing paper describes an outlook it calls more software engineering oriented than other LLM agent approaches: AutoCodeRover works on a program representation, the abstract syntax tree, rather than viewing a software project as a mere collection of files. Its code search exploits program structure in the form of classes and methods, which the authors say enhances the LLM’s understanding of an issue’s root cause and effectively retrieves context via iterative search. Spectrum-based fault localization sharpens that context further, on the stated condition that a test suite is available.
The same paper reports its measured behaviour on SWE-bench-lite, described there as 300 real-life GitHub issues: 19% efficacy in solving them, which the authors state is higher than the recently reported efficacy of SWE-agent, at an average cost of $0.43 USD — significantly lower, in the authors’ description, than other baselines.
Adoption & Ecosystem
A survey on AI agentic programming classifies AutoCodeRover, in its comparative taxonomy, as a “Multi-agent System” that is proactive, multi-turn, tool-using, and adaptive.