Jev 单次 agent 运行做 973 个决策、零输出 token
同一 agent 流程改用 Jev 做路由、打分与工具调用安全门后，成本从 $22.18 降到 $0.05，回答 38ms 返回，前沿模型一次都没被调用。
Jev made 973 decisions in one agent run and generated 0 output tokens doing it same run, same output: $22.18 on a reasoning model before, $0.05 after, and the frontier model was never asked once the loop: prompt → structured state → Jev picks the model → worker runs → Jev scores the result Jev gates the tool call deliverable each decision ships a small state object, not the transcript. that is why the input side stays cheap and the answer comes back in 38ms instead of seconds each decision ships a small state object, not the transcript. that is why the input side stays cheap and the answer comes back in 38ms instead of seconds the router: 7 models on the table, Jev picks one per step. 44% went to the cheapest, 0 calls went to the frontier one the gate: every bash, browser and write call gets p(safe) before it fires. 39 verdicts, 1 held, 0 approvals asked of me the filter: 1,204 sources in, 214 dropped, kept at 0.81 a scored decision returns probabilities over the options, so there is nothing to generate. mean confidence 0.91 the model still writes. it just stopped deciding full breakdown below ↓
这条是引用着另一条写的：x.com/0xcarnagee/status/2101019382009004249