把 agent 里的小决策从主模型挪到 Jev 的工程做法
作者建议审计 agent 会话中只为说「是/否」而写的句子，把工具调用前放行、模型档位选择、任务是否真完成这三类判断交给 Jev，半秒内返回，一万次决策约 0.42 美元，走主模型则约 300 美元，并附一段可直接粘贴的审计提示词。
Jev engineering tip: open your agent's last session and count how many times it wrote a sentence just to say yes or no put the small decisions somewhere that can't write a sentence Your agent keeps writing, editing, researching. Jev sits underneath it and only gets involved at three points: → before a tool call runs: safe, or does a human need to see this first? → when the task needs a model: the cheap one, or the expensive one? → before marking done: did the test actually pass, or does it just look finished? Jev answers in under half a second, and that speed is why 10,000 of these decisions cost 42 cents through it. route them through your main model instead, and the same 10,000 cost roughly $300 the full loop your agent runs the main session Jev checks every tool call before it executes Jev routes the task to the right model anything under the confidence threshold escalates back up Most agents never audit which fork is which. They just pay full price for both. paste this into your coding agent: "Audit my agent's decision points before we change anything. > List every model call in the last real run. For each one: what triggers it, what the output gets used for. > Classify each call: could every valid answer be listed in advance, would a person settle it on a glance, does it happen often enough to matter. Mark GENERATION or DECISION. > For every DECISION call, show what it would look like as a bounded decision instead, plus a confidence threshold below which it escalates back to the main model. > Report which calls stay, which move, and the real cost difference at current volume. If there's not much to gain, say that plainly. Show the classification as a table first. No changes until I say go."
这条是引用着另一条写的：x.com/rvaniaaaa/status/2101715726398579131