Jev as a cheap decision layer in front of agents
The author argues that using Jev for thousands of small YES/NO decisions in an agent frees expensive models for research and planning, and lists 12 integration points and routing configs.
whoever built this noticed we keep hiring the smartest model for the most boring work you pay $25-$250/mo for a brain that can research, code and plan… and a big chunk of its day goes to answering YES / NO. Jev takes that job off its plate. it's a small decision model that sits in front of the expensive calls and answers the boring forks: -> which agent runs next? -> worth researching or skip? -> reply or ignore? -> publish, wait or escalate? -> BUY / SELL / HOLD? -> click or keep looking? the reported numbers are what made the setup click for me. ~2,000 papers sorted for about $0.05 ~800 emails triaged for about $0.028 a browser-agent loop finished in ~6 seconds for ~$0.0029 and at $0.042 per million input tokens with free output, ~20,000 short decisions (~700 tokens each) lands near $0.62. Jev isn't out-thinking Claude here. it was never asked to think. the split is almost boring: -> big LLM does the research, writing and planning -> Jev makes the thousands of small calls in between -> code does the clicking, saving, sending and executing once you see it, the forks are everywhere. > inbox triage. > X replies. > support tickets. > meeting notes. > lead scoring. > clip picking. > browser agents. > even paper-trading loops. this is the part most agent setups get wrong. a frontier model shouldn't sit on every step just because it can. > Claude does the thinking. > Jev picks what's worth Claude's time. > code does the work. I mapped 12 spots to wire this into an agent today + the exact router config in the article below
This post quotes another post: x.com/asteri_eth/status/2104901781725598103