夜间值守 agent：Jev 先筛，前沿模型只处理 0.24%

Superior @andreysuperior原推发布于 2026.10.06

在 X 上打开原推

A model that can't write a single word now runs the whole AI night shift for $2.49. Your always-on agent spends the night asking itself tiny questions, and right now it pays frontier prices for every one of them. Same night with everything on GPT-6 Astra: $403. Same night with Jev deciding first: $2.49. Here is who should answer each 3AM question: 1. "Is this email worth reading?" Jev. A yes or no with a confidence score, for a fraction of a cent. 2. "Which worker takes it?" Jev picks one and tells you how sure it is. 3. "Retry or move on?" Jev, with a hard step limit in code so the loop can't spin all night. 4. "Does this need a reply?" Jev says yes, GPT-6.1 Sol writes it at a fifth of Astra's price. 5. "Is this 90-page contract hard?" Jev rates it, and only then does GPT-6 Astra wake up. 6. "Delete the customer? Send the refund?" Never the model. Code checks the rules and it waits for your click in the morning. > On the board below: 39,340 decisions in one night, and the frontier model ran on 0.24% of them. The full night-shift build is in the article below:

这条是引用着另一条写的：x.com/noisyb0y1/status/2105964235087856080