Replacing 85% of an agent's model calls with Jev decisions
In a typical 13-call loop, 11 calls are just picking workers, scoring, deduplicating and gating, so the author hands ROUTE/KEEP/LOOP/GATE decisions to Jev and leaves Opus only the two calls that actually need writing.
85% OF YOUR AGENT'S MODEL CALLS DON'T NEED A MODEL THAT CAN WRITE. YOU'RE PAYING OPUS PRICES FOR EVERY SINGLE ONE. A typical loop: 13 model calls. 11 are decisions. 2 are real generation. Which worker next? Safe to run? Still relevant? Stuck? Done? Yes, no, pick one, rate it. Not one word needs writing. ROUTE - Jev picks the next worker KEEP - Jev scores which context survives LOOP - Jev catches the agent repeating itself GATE - Jev blocks risky tool calls, uncertainty goes to review Opus keeps the 2 calls that matter: planning, writing, coding, synthesis. The decision side costs almost nothing: 70–500 ms per call $0.042 per million input tokens, output free 10,000 decisions ≈ 42 cents One Claude session went from ~1M tokens to 86K in one second. Nothing summarized. Just scored and dropped. The smart model was never the problem. Asking it yes or no was. Save this before your next agent build.