用 Jev 做决策层替 agent 省下前沿模型调用

morbian @morbianai原推发布于 2026.10.22

在 X 上打开原推

Your AI agent is burning $100s a month answering yes/no questions. Here's the fix. every time your agent checks "is this spam?" it calls a frontier model. every "should I reply to this?" - frontier model. every "is this click safe?" - frontier model. that's like hiring a senior engineer to sort your mail. so I built a different setup. one rule: LLM writes. JEV decides. code acts. JEV is a small, cheap decision layer that sits in front of your LLM. it handles the boring calls with typed outputs and a confidence score, and wakes up the big model only when there's something to actually write. here's what it routes in the demo: -> email triage phishing=0.97 → quarantine(). no LLM touched it -> youtube comments worth_reply=yes → llm.draft(). the model only writes replies that matter -> inbound leads lead_quality=high → calendar.send() -> browser agent picks the next action before your agent clicks blindly -> x feed viral_fit + timing → queue.reply() -> irreversible actions safe_to_run=no → gate.hold(). a human signs off the numbers on screen: 127k decisions a day. ~248ms. $0.42/day. here's the uncomfortable part: your agent isn't dumb. it's just expensive in the wrong places. stop paying GPT prices for a boolean. full breakdown with use cases below