jev-kit：Claude 写代码、Jev 做判断的浏览器 agent

Deniss @learnodd原推发布于 2026.10.17

在 X 上打开原推

I read the jev-kit browser table three times. I was sure the decimal point was in the wrong place someone wired Claude Code and TypeSafe's Jev into one team, and the agent stopped paying a frontier model for every yes/no most people run agents in one of two ways: -> let one big model decide every click, every tool call and every sub-agent, and pay frontier prices for forks a cheap judge could name -> hardcode if/else rules that break the first time the task changes jev-kit splits the stack in two: > the writer: Claude writes the code, the drafts and the text in form fields > the judge: Jev returns a typed choice, score or yes/no with a confidence, in about 0.3 s warm > the guard: every tool call hits a rules table first, only the ambiguous ones reach Jev > the tier check: a sub-agent sent 2+ rungs too expensive gets blocked, cheaper type named > the belay: an agent that says "done" with no passing check gets sent back to verify and the numbers, all measured in the repo: > 9/9 browser goals for $0.0008 of Claude per run vs $0.1868 with Sonnet deciding > 98.2% tier-guard accuracy on 57 scored dispatches, zero false blocks > 33.2 ms added per tool call when no rule matches > 100% on the deny-capable rules eval, zero false denies Claude writes. Jev decides. run it in shadow mode for a week, read the log, then let the cheap judge take the calls

这条是引用着另一条写的：x.com/hanakoxbt/status/2110067711363457244