Jevgrep Reduces Coding Agents' Need to Search for Code Independently: SWE-bench Main Model Costs Drop by 29%

Beating AI Insight News Brief: Developer David has open-sourced Jevgrep, a research tool built on Jev specifically for coding agents to locate code. Users only need to ask a question like "Where is the login validation implemented?" and it uses TypeSafe’s Jev decision model to search the codebase layer by layer, identifying relevant files and source code snippets before passing them to agents such as Claude Code and Codex for further modification and testing. It primarily addresses the token-intensive "code lookup" step that coding agents face. Jevgrep does not feed the entire repository to the model upfront, nor does it rely solely on a single semantic search. It first determines which directories are worth exploring further, then checks relevant files and code declarations, and finally returns source code snippets, line numbers, and follow-up reading references. The repository also includes a "Skill" that enables agents to 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 the total cost for GPT-5.6 Sol dropped from $7.62 to $5.44, a reduction of 28.63%. The author initially stated a 40% reduction on X, but later updated the experiment results in the repository, revising it to "around 30%". However, this figure only accounts for Sol’s costs and does not include Jev’s expenses. The confirmed Jev call costs in the experiment logs amount to at least $1.57, while some calls lack complete billing records, so the actual total is unknown. Additionally, the experiment only included 10 tasks, with each run only once.