Jev engineering runs a 40-agent research system on a single laptop
Jev moves judgment from the large model to a small model that answers in types (Choice, Score, probability), handling 40 questions in parallel with sub-second returns while the large model only writes.
JEV ENGINEERING TURNS A 40-AGENT RESEARCH SYSTEM INTO SOMETHING THAT RUNS ON A SINGLE LAPTOP Most agent stacks send every judgment back to the language model, so each small decision costs a full call and the bill grows with how careful the system is. The deciding moves out to a tiny model that answers in types instead of sentences, a choice, a score, or the odds something is true. One call carries 40 questions in parallel, scope checks, risk scores, retry ceilings, all landing back in under half a second. The big model stops seeing decisions entirely and only writes, because judging and generating turned out to be two jobs on two different price tiers. Running six projects at once barely moves the cost, since the part that scales is now the cheap part. See how the typed question layer actually works below👇
This post quotes another post: x.com/0xcodio/status/2104195432209387524