600 个 Claude agent 用 Jev 做置信度闸门
600 个 Claude agent 分 30 组跑 5 类任务，Jev 给每个输出打置信分，低于 0.90 的拦下人工复核，4 个自动放行、1 个被扣下。
600 CLAUDE AGENTS, ONE CONTROL CENTER, ZERO UNCHECKED OUTPUTS Speed isn't the impressive part. The impressive part is that nothing goes out unverified I opened a real-time control center: 600 Claude agents split into 30 squads, with Jev running backend checks on every output. Each result gets a confidence score, and anything under 0.90 lands on my desk instead of going live I gave them 5 jobs at once Sales lead: buying signal 0.93, demo booked Pull request: 214/214 tests green, merged at 0.97 Invoice: PO matched at 0.98, payout queued Churn alert: scored 0.86, held for my review Bug report: P1 at 0.95, on-call paged Four went through on their own. One waited for me. That's the whole idea $2 spend cap, 12 actions max per job, and a human signs off on anything uncertain This isn't just another agent show-off. It's a system that knows when it's unsure and stops before it breaks something If you know how to build this, you will never be out of a job If you don't, the article below is made specifically for you
这条是引用着另一条写的：x.com/genuenci/status/2106475874333675862