Dots 与 Jev 分工的 6 步蓝图
拆解 Dots 用规则分派任务、Jev 用 0–1 概率在 70–500ms 内判定风险并拦截 prod 迁移，附审计链路。
holy sh*t, Sam Altman launched Dots, but Jev is still the one that decides the fork Dots does the work while you're away, and Jev makes sure every decision inside your own agent is one you can audit this is a 6-step blueprint on how Dots and Jev each handle a fork, and when you need which: step 1 → Dots starts from a goal: you set it once, like "weekly market brief, ask before email", and your dot on GPT-6 Astra runs on its own cloud computer with memory that stays the same in every channel step 2 → Dots splits the work: background agents research the web and Drive, analyze the numbers and draft the PR in parallel, each one a thread you can open step 3 → Dots decides with rules: every action passes an automatic review of instructions, permissions, custom rules and safety, then lands in one of four modes: without asking, when you say so, ask first, or hand off to you step 4 → Jev starts from the state: "the agent wants to run a migration on prod-db" goes in as one packed request step 5 → Jev decides with numbers: typed questions like "is prod touched?" and "how risky, 0 to 1?" come back as calibrated probabilities in 70–500 ms. risky 0.93 is above the 0.90 deny line, so code blocks it step 6 → Jev hands the next step back to the LLM: it writes a safer plan, a dry-run on staging, and the result is verified in code, leaving an audit trail for every fork most people think these two compete they don't: Dots is the agent you use, and Jev is the layer you build into your own use Dots when you want the work done while you're away, and use Jev when you're building the agent and need forks you can audit ↓
这条是引用着另一条写的：x.com/gyome1_/status/2098057823930155319