Jev + Claude Code = The Cheapest Agentic Coding Loop Yet
Ray Amjad explains TypeSafe's Jev model as a fast System 1 probability classifier. He demonstrates its three primitives—Noul, choice, and score—in the web console before exploring agentic patterns like parallel browser testing, dynamic skill selection, qualitative linting, and multi-angle PR review paired with System 2 frontier models.
Original by Ray AmjadAgent workflowsIntermediate27 min 28 secPublished
Use a yes-or-no proposition, a supplied option set, or an ordered scoring rubric instead of asking Jev to write review prose.
Ray describes a loop where a generative model improves evaluation criteria and Jev runs the bounded checks.
The examples include skill selection, qualitative linting, and pull-request screening before more expensive model review.
Worth knowing
Fast latency and high probability scores do not guarantee semantic accuracy or catch every category of bug; community benchmarks and cost estimates reported from tweets are unverified anecdotal claims.