David shipped jevgrep, claims 40% lower coding agent cost
A CLI that gathers context with Jev before the coding agent starts spending tokens on it.
David released jevgrep, a research agent CLI powered by Jev that he says reduces coding agent cost by 40%, verified on SWE-bench. The pitch is a split of labour. Jev does the context collection, the expensive coding model only sees what survives. The tool ships with a built in skill so the coding agent knows to call jg when it needs context rather than going and reading the repo itself.
a research agent CLI powered by jev from @typesafeai that reduces your coding agent cost by 40% (verified on SWE-bench)
That is the whole shape of it, and it is the shape a lot of people have been reaching for. Context gathering is the part of an agent run where token spend piles up fastest and where the reasoning required per step is lowest. Grepping, opening files, deciding whether a match matters. If a cheap fast model can make those calls, the costly one starts from a smaller and better pile.
What is not in the post
The 40% is David's own number. There is no baseline model named, no task subset listed, no before and after token counts, and no cost per instance. "Verified on SWE-bench" tells you the benchmark and nothing about the configuration, and SWE-bench results move a lot with the harness. Until somebody else runs it against their own agent and posts the two numbers, treat it as a claim from the person who built the thing, not as a result.
The reception so far is one line. Theo replied that it is an interesting use case, which is notable mostly because he has been the sharpest public skeptic on this exact boundary, having argued earlier that Jev cannot validate because it cannot run tools. jevgrep does not contradict that. The CLI is the thing holding the tools, Jev is the thing deciding what comes back.
For people building on it, the number to watch when you try this is not cost alone. A context filter that is cheap and wrong costs you a failed run, and a failed run is more expensive than the tokens you saved. Measure resolve rate next to spend, on your own repos, before you wire jg into anything that ships.
ooo this is actually an interesting use case