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

Modern LLM coding agents like Claude Code and OpenHands often waste token budgets on repository exploration. CodeGrep, a 14B retrieval agent trained with GRPO, optimizes this by issuing parallel grep, glob, and read calls. On SWE-Bench Verified, CodeGrep increased the resolve rate to 27.0% from 25.8% while improving efficiency. This complements the standard agent loop of Look, Think, and Do, which allows agents to iteratively analyze context, plan changes, and verify fixes in complex codebases.