Jevgrep helps Coding Agent find less code on its own: SWE-bench main model cost drops 29%

动察 Beating AI News Flash: Developer David has open-sourced Jevgrep, a research tool built on Jev specifically designed to find code for Coding Agents.

Users simply ask "Where is the login validation written?" and it uses TypeSafe's decision model Jev to search the codebase layer by layer, identify relevant files and source code snippets, then hand them off to Agents like Claude Code and Codex for further modification and testing.

It primarily addresses the token-expensive "code finding" step in Coding Agents. Jevgrep doesn't stuff the entire repository into the model upfront, nor does it just do a single semantic search. It first determines which directories are worth continuing to search, then examines relevant files and code declarations, and finally returns source code snippets, line numbers, and follow-up reading leads. The repository also provides a Skill that lets Agents know when to call `jg` to gather context.

The latest SWE-bench experiment used 10 Python tasks. Both with and without Jevgrep, 8 tasks were completed, but GPT-5.6 Sol's total cost dropped from $7.62 to $5.44, a reduction of 28.63%. The author initially wrote 40% on X, and the repository subsequently updated the experimental results, now changed to "approximately 30%."

However, this figure only accounts for Sol's costs and does not include Jev. The experiment logs confirm at least $1.57 in Jev call costs, with some calls lacking complete billing records, so the actual total is unknown. Additionally, this experiment set only had 10 tasks, and each task was run only once.