用 Jev 做 agent 行为监控省 126 倍成本

Ridark @ridark_eth原推发布于 2026.10.02

在 X 上打开原推

monitoring ai agent actions with frontier llms cost $372. a classifier called jev did it for $2.94. i read that twice. 126x cheaper, and the big models look absurdly overpriced next to it. the reason is what jev is allowed to output. agent action → jev → probability per option → block / flag / allow → $2.94 bill → every agent action gets checked against its assigned task → you write the options, up to 255 per question. typesafe's jev returns a probability for each → confident it's a problem, blocked. uncertain goes to a human, the rest pass → list price is $0.042 per million input tokens, output free → the bill is under 1% of the llm one. a dollar check there costs under a cent here → the $372 that paid for one run covers 126 runs → openai showed its own version at devday on sept 29, on its small luna model. limited preview, no pricing one caveat. both prices come from one querystory demo, workload size not stated. jev's output is one score per option you wrote. it never writes an answer. watch one action go through the gate ↓

这条是引用着另一条写的：x.com/ridark_eth/status/2084625062749958537