Routing agent tasks with typesafe jev to cut costs
Jev labels each task fast or think and routes it by threshold, sending only 17% to the expensive model and cutting monthly cost from $64.6k to $44.7k.
the smartest 24/7 ai agent isn't a bigger model. it's a f**king router that knows when NOT to think... running openai dots on gpt-6 astra for every incoming task is insane: → 1,086 tasks/session, $0.059 avg if everything thinks → $64.6k/mo for ONE agent that resolves email and tickets here's what i'd actually build instead: [the triage line] → typesafe jev reads every task and stamps it: fast p=0.91 or think p=0.72 → threshold at ~0.29 decides which lane the task falls into → fast lane: $0.07/task, 0.36s latency, 86% resolved on its own → think lane: astra only wakes up for the 17% that need a brain → confusion matrix catches misroutes before they hit a human the kicker: route accuracy stays at 85.4%. miss rate under 10%. nothing risky gets silently closed, everything above threshold hands off to astra. result: $64.6k/mo → $44.7k/mo. −52% burn. same inbox cleared. you don't need a smarter model. you need an agent that knows when to think. full triage-line playbook in the article below ↓
This post quotes another post: x.com/shmidtqq/status/2108637104016072888