Jev 编码循环：32 步 158 次决策的实测拆解
作者复盘一次 Jev 编码 agent 运行：按决策而非任务路由模型，158 次判断成本 0.0157 美元，并给出 halt gate 与破坏性阈值等 5 条设计经验。
JEV DOESN'T WRITE YOUR CODE. IT MAKES THE FIVE DECISIONS AROUND EVERY LINE OF IT. HERE'S THE LOOP every step your agent takes is really five judgment calls. which files to read. which model writes this. is this safe to run. keep this output or drop it. are we actually done. today one frontier model makes all five, and you pay frontier prices for every one. jev coding loop: 32 steps, 158 decisions, 7 things shipped, 3 sent back to the human. i went through the whole run. 5 things before you copy this design. one station writes the code. jev sits around it and scores: files by relevance, the write by difficulty, the command by blast radius, the output by whether it earns its place, the task by whether it's done. -> ROUTE by decision, not by task. this run went cheap 48%, mid 37%, frontier 15%. most calls in a coding loop are small judgments, not hard reasoning. the frontier model shows up for the 15% that needs it. -> COUNT the cost both ways. 158 decisions cost $0.0157 through jev against $20.46 through a frontier model. but a judge that runs five times per step makes a lot of calls, so cheap per call is the only version of this that works. -> WATCH the halt gate. 153 allowed, 3 sent to the human, 2 halted and asked. that's the whole safety story: the loop stops itself instead of asking forgiveness after the delete. -> CHECK the destructive threshold. one line in the log reads "destructive 0.78 held." a number decided your file didn't get wiped. set that number yourself, don't inherit it. -> READ the footer. it says simulation, illustrative values, rates as published sep 2026. this is the shape of the system, not a benchmark. treat the dollar figure as a hypothesis until you run it on your own repo. the idea is right: stop asking the biggest model to make the small calls. a judge that costs a fraction of a cent can stand next to every step, and the expensive model only gets woken up when the step is actually hard.
这条是引用着另一条写的：x.com/me_barnyx/status/2101976764779999381