OpenAI Dots 一周托管的分权与检查点做法
作者把 Dots 拆成四角色、限制链长 5 步、权限分三档、每日审批不超 5 次，主张给自主性加硬边界。
openai dots looks simple until you let one agent run your whole week i turned the setup into a tighter system instead of giving one dot unlimited freedom: step 1 → give it 4 seats: LOOKOUT, MAKER, SKEPTIC, RUNNER. one agent, one memory, one role at a time. step 2 → never let a chain run past 5 steps. stop, run a checkpoint, then continue. step 3 → split permissions into allow / ask first / block so uncertainty always moves toward you, not toward action. step 4 → keep approvals under 5 a day. if you’re clicking 40 prompts, the guardrail already failed. step 5 → give every artifact somewhere permanent to land, then keep a simple index of what the agent actually did. the useful shift isn’t “make the agent more autonomous.” it’s giving autonomy hard edges. full OpenAI Dots guide + prompts in the article below ↓
这条是引用着另一条写的：x.com/fleyta88/status/2105656875387625507