四个 agent 值夜班：Jev 做廉价决策层
ANT/PAUL/ZERO/CORTEX/FLY 分工跑一夜，用 Jev 每百万次决策 $0.042 处理例行判断，贵模型只留给难例，ZERO 在删客户前自动刹车。
4 AI agents ran the night shift for $42 per million decisions, and one of them refused to delete a customer. Here's what the log said by morning: 18 pairs scanned. 3 thin books flagged. 12 shops set open. 42 links written to the vault. 1 customer delete put on HOLD. Every agent has one job. ANT watches the markets and flags thin liquidity. PAUL opens the shops and queues restock notes. ZERO drafts customer replies. CORTEX turns everything into linked notes. FLY carries packets between them. Look closely: almost nothing here is hard. "Is this book thin?" "Is this shop open?" "Does this need a reply?" Small decisions, thousands of times a night. Run all of it on a frontier model and the bill kills the project in a month. So you build a ladder: A cheap decision model for routine calls, something like TypeSafe's Jev at $0.042 per million. A mid-tier model for drafting, GPT-6.1 Sol. The frontier model, GPT-6 Astra, only for the genuinely hard cases. The best moment in the video is ZERO. It drafts the reply, then stops itself before delete_customer. That's the rule for any agent that runs while you sleep: Cheap model decides. Code authorizes. Humans approve what can't be undone. CORTEX is the quiet winner. 42 links in one night means tomorrow's agents start smarter than today's. Your morning: you don't do the work, you read the log and approve what's on HOLD. Most people give one giant model the whole night shift and wake up to a huge bill. A few build a team with rules.
这条是引用着另一条写的：x.com/noisyb0y1/status/2105964235087856080