Jev 当 agent 的 System 1 决策层省下 98% 调用
演示把 agent 拆成 System 2 推理与 Jev 快速选择两层,1 万次决策中 7100 次不再走大模型生成。
Jev Engineering is the System 1 most agent stacks are still missing. same agent → two speeds → one split $98.41 pulled out of 10,000 decisions. everyone is optimizing the model. Jev Engineering optimizes what never needed a model. every agent eventually hits the same fork: state → draft / pick / score / yes-no / code → one lane has to win instead of sending every “which worker?” and “how urgent?” back to a $15 / M call, Jev turns the reflex into its own layer: so the loop becomes: System 2 → research, write, reason Jev → pick, score, gate code → execute state → update Jev → decide again once that split is real, 7,100 of the 10,000 calls leave System 2 with 0 gen tokens.
这条是引用着另一条写的:「长文:Jev 或成 AI agent 缺失的控制层」