The author's solution separates thinking from doing: orchestrators only write the task, and the small model Jev decides which model (Haiku, Sonnet, or Opus) executes it. It scored 40/40 in testing, and currently only about 1 in 18 tasks goes to Opus.
A saída foi separar quem pensa de quem faz. Orquestradores só escrevem a tarefa. Um modelo pequeno, o Jev, decide quem executa e com qual modelo: Haiku, Sonnet ou Opus. Acertou 40 de 40 no teste. Hoje só 1 em 18 tarefas vai pro Opus.
LLM holds memory/context, uses agent-desktop to take snapshots and sends them to Jev, which selects the right element to act on for faster Computer Use, with Cursor enabled and headless support.
Riley Brown shares that TypeSafe AI's Jev classified 500 emails in seconds for just 3.5 cents, highlighting its fast and low-cost email classification.
A comprehensive explainer video introducing Jev, an AI that doesn't write text but determines which department to route inquiries to and how urgent they are.
