实测把 agent 决策层换成 Jev 后成本从 850 美元降到 0.2 美元
作者七天实测：Claude 只负责起草，Jev 从代码生成的选项里做选择，代码保留最终决定权，千次决策成本从数百美元降到几美分。
JEV + FRONTIER LLM = 10X LOWER COSTS I tested my agentic system to see if and where I can implement Jev After I finished 7 days test, the results shocked me: $850 in tokens spent on decision making, when with Jev it would cost me <$0.20 So I upgraded my system: > Claude drafts. > A decide layer picks from a code-built menu. > Code keeps the final say. > Every step writes a receipt. In the writeup they clocked ~1,000 cold frontier decisions at a mid-hundreds bill. Same thousand through the harness lands in the low-cents range. How the spend dies: ➔ Plans and patches only from the big model ➔ Illegal routes cut in code before any pick ➔ Decide layer answers or abstains ➔ Live state revalidates, then the tool runs under your normal approval Steal without copying their stack: ➔ One checkable decision per step ➔ Options prepared in code, not chat vibes ➔ A written path for "not sure" ➔ Logs you can replay when policy changes The engineer who can show that collapse in a room designs the next system. Model writes. Layer chooses. Code executes.
这条是引用着另一条写的：x.com/kaiandai/status/2103871840066130241