把 agent 的小决策交给 Jev：100 封邮件 1.42 秒 7 美分

Lucky @Luckyjlg5ir原推发布于 2026.09.22

在 X 上打开原推

This developer connected a model at $0.042/M tokens and discovered what his agent had been overpaying for all along Jev from TypeSafe AI doesn’t write code, do research, or generate answers. It reads the system state and chooses from predefined options: which worker to run, which source to keep, whether a review is needed, or whether it can move forward. The output is a typed decision with a confidence score. 70–500 milliseconds per decision. In one test, 100 emails took 1.42 seconds, with 96 correct answers and a total cost of seven cents. Then 500 emails for 3.5 cents, while a Claude session was compressed from nearly one million tokens to 86,000 in just one second. The tasks are different, but the underlying operation is the same: the system doesn’t need to write new text, it needs to choose one option from several. That’s exactly the kind of small decision an agent makes between every major action. Which worker to run, which source to discard, whether there is enough data, whether a review is needed, whether an action can be executed. If every one of these choices goes through a full LLM, you’re paying for generation where the result will still be just one short decision. If an agent has a lot of these calls, they can be moved into a separate decision layer. The LLM stays responsible for code, research, and generation, Jev handles the choices, and code executes the decision. The agent doesn’t become magically smarter - it simply stops using generation where all it needed was to choose the next step.

这条是引用着另一条写的：x.com/danikkk_wqs/status/2101706326883655710