Coding agents spend 631K tokens hunting for the right file

An OpenHands coding agent burns 631,000 tokens and 23 rounds resolving a single SWE-Bench Verified issue. Most of that goes to grepping, globbing and viewing files, per the CodeGrep paper.

That's the compute bill for finding the code, before the agent writes a line of the actual fix. File search is the cost center, not reasoning.

Researchers built a 14 billion parameter retrieval agent trained end to end with GRPO and dropped it into the same loop. Resolve rate on SWE-Bench Verified climbed to 27.0% from 25.8%, while rounds fell 15% and tokens dropped 19% on resolved instances.

Precision has a floor

They also ran off-the-shelf retrievers through the same test, and the results split hard on one number: precision.

BM25 measured 0.375 precision and made the agent's results worse, not just slower. Jina hit 0.445 and came out neutral, no gain, no loss. CodeGrep, at 0.677 precision, is the only one that helped the agent resolve more issues.

The paper tests one benchmark and one agent size. It doesn't say whether that precision line holds for larger models or codebases outside SWE-Bench Verified.