CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents

A recent paper published on arXiv (2608.05886) presents CodeGrep, a 14B retrieval agent developed end-to-end with GRPO, aimed at enhancing the performance of LLM coding agents such as Claude Code and OpenHands. These coding agents typically expend excessive tokens on exploring repositories, averaging 631K tokens and 23 rounds for each resolved issue on SWE-Bench Verified. CodeGrep employs multi-turn parallel grep, glob, and read tool calls to provide candidate files to a static downstream coding agent. Across all 500 instances of SWE-Bench Verified, CodeGrep maintains a resolve rate of 27.0%, surpassing the 25.8% baseline, while achieving a 15% reduction in rounds and a 19% decrease in tokens for resolved cases. The paper also evaluates retrieval precision thresholds, indicating that BM25 with a precision of 0.375 negatively impacts the agent, whereas Jina with a precision of 0.445 remains neutral. This research was introduced as a cross-type submission on arXiv.