Layered agent setup: Jev routes in ~16ms while Opus only handles architecture
The author sketches a four-tier agent split: Haiku scans the repo, Jev routes routine work in about 16ms, Sonnet writes code, and Opus wakes only for plans, repeated test failures, or pre-ship checks.
STOP USING OPUS FOR EVERYTHING. YOU'RE PAYING A SENIOR ARCHITECT TO RUN GIT STATUS I did it for months. Opus on every prompt. Then I saw how the good setups actually work, and drew it as a brain: > HAIKU swarms the repo. file discovery, AST scans, grep, docs > JEV routes the boring stuff in ~16 ms. no frontier model wakes up > SONNET runs the main session and writes the code > OPUS sleeps. on call. the architect nobody bothers Opus gets woken for exactly three things: -> a plan big enough to hurt if it's wrong -> the same test failing twice -> the moment before you ship lint, git status, rg TODO? red X. never reaches it. You're not saving Opus money. You're saving Opus attention.
This post quotes another post: x.com/loofyb0i/status/2102452178384273431