Jev 只做决策、Kimi 只写文本的分工省钱法
把 agent 里 91% 的「是/否、选源、打分」类调用交给 Jev 返回概率，写作交给 Kimi K3，决策成本降约 250 倍。
SAM ALTMAN AND DARIO AMODEI CHARGE YOU WRITER PRICES FOR A YES OR NO. 91% of my agent's calls were exactly that. I moved them. The bill fell 11x. 1 week of logs: - 1,100 calls - 1,003 decisions - 97 texts I kept Every decision went through the big writing model. At writing prices. So every bot got one job: - Jev decides - Kimi K3 writes - plain code does the rest A decision is tiny. Pick a source. Yes or no. A score from 0 to 1. Jev returns a number with a probability, code acts. Nothing written, nothing billed as writing. Jev's 4 jobs: 1. Route → picks the next worker. Above 0.85, no second opinion 2. Escalate → below 0.85, Kimi K3 takes the case 3. Gate → under 0.95, I say yes or no 4. Compact → drops useless history before Kimi K3 reads it One brief: - decisions: $0.42 → $0.0017 (~250x) - writing: $0.43 → $0.071 (6x) 100 briefs a month: $85 → $7.30 One agent, one week. Your split will move with your traffic. Steal this: Log a week. Count choices vs kept text. Send the choices to a cheap gate. One-person desk. Named bots, one job each. Ready to copy. Save this.
这条是引用着另一条写的：x.com/starmexxx/status/2108159970017488953