NetEase's CodeGrep: An RL Agent Cutting Coding Agent Token Cost by 19%
_reachsumit · x · 2026-08-07
Modern LLM coding agents like Claude Code and OpenHands waste much of their token budget finding files to patch. On SWE-Bench Verified, a 30B OpenHands agent averages 23 rounds and 631K tokens per resolved issue.
To solve this inefficiency, researchers introduced CodeGrep, a 14B retrieval agent. Trained end-to-end with GRPO, it issues multi-turn parallel grep, glob, and read tool calls to return candidate files to a frozen downstream coding agent.
Across all 500 SWE-Bench Verified instances, CodeGrep preserves the resolve rate (27.0%) while substantially improving efficiency: it reduces rounds by 15% and tokens by 19% on resolved instances. The study also identifies a precision threshold for retrievers, showing that retrieval only reduces rollout cost when precision is sufficiently high.
More from coding & agent
- Inverted Task Synthesis: 37K Long-Horizon Tasks at $0.05 Each, DeepSeek Pass Rate Drops to 2.5% — teortaxesTex · 2026-08-07
- Aeon: A Fully Autonomous Agent Framework Without Approval Loops — tom_doerr · 2026-08-07
- Loomline: AI Platform Aims for Full-Lifecycle Software Delivery Beyond Code Generation — jiqizhixin · 2026-08-07
- Stop Making Up Names: YacineMTB Argues Good Models Just Need Bash for Agents — yacineMTB · 2026-08-07
- Open-Source AI Video Ad Studio Turns Product Photos into Full Ads — Fresh-Resolution182 · 2026-08-07
- DeepSeek+Luna Cascade: 78.9% Accuracy at 63% Cost, Beats Luna Alone — zainhas · 2026-08-07