[research] Coding agents waste 19% of tokens just finding files to edit

[github-actions[bot]]

🔬 The Finding

Researchers at arXiv found that LLM coding agents like Claude Code and OpenHands spend a disproportionate share of their token budget on repository exploration — grep, glob, and file reads — rather than actually solving problems. On SWE-Bench Verified, a 30B OpenHands agent averages 631K tokens and 23 rounds per resolved issue, much of it just locating the right file. Their solution, CodeGrep (a 14B RL-trained retrieval agent), prepends a parallel file-finding step that cuts 15% of rounds and 19% of tokens on resolved instances while slightly improving resolve rate (27.0% vs 25.8%).

⚙️ What It Means for Agentic Workflows

1. Specialize your retrieval step. Don't let your primary agent burn context budget exploring a repo. A cheap, dedicated retrieval layer (even a small model) can front-load file discovery and hand the relevant files to the main agent, dramatically reducing costly round-trips.
2. Parallelize tool calls during exploration. CodeGrep issues multi-turn parallel grep/glob/read calls — a pattern directly applicable to any GitHub workflow agent that needs to navigate large codebases before acting.

🔗 Source

CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents — August 6, 2026

Generated by Daily Agentic AI Research Digest · 97.2 AIC · ⌖ 10.5 AIC · ⊞ 24.2K · ◷

• expires on Aug 16, 2026, 9:54 AM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-08-16T09:54:13.747Z.

Closed by Workflow