The Rule Was Only a Wish
Your agent fixed the deck overnight. Nobody asked it to.
Remember
Code and permissions hold. Rules are requests.
Context
Dots launched at OpenAI DevDay on 29 September 2026: persistent personal agents in ChatGPT, each a named mascot with its own cloud computer. This Short is the 2:30 companion to the landscape deep dive ai-visual-deep-dive/specs/chatgpt-dots-deep-dive.md (8:34). Owner's decision: no villain — the Short explains the mechanism; the deep dive keeps the two failure stories.
Architecture
| Component | Role | Notes |
|---|---|---|
| Triggers | Wake the dot without a prompt | Morning timer, new email, Slack mention |
| Proactive research | Reads connected apps while you're away | Read-only, enforced in code; output is a suggestion or question |
| Action tier | Does the work once you say yes | Own cloud computer (browser, files persist); apps via plugins |
| Private notes | What the dot learned | Can't be viewed or edited one by one; deleting the dot is the only reset |
| Custom Rules | Your preference per action | Four levels; words the dot reads and tries to follow |
| Auto-review | Second model, outside the dot's control | Checks every consequential action |
| App permissions | Per-app level | Always ask · Read-only actions · Allow low-risk · Allow all |
| Fixed floor | Hard limits in code | Password changes, money transfers always to you; deletes/installs approved every time |
| Activity View | Audit | Desktop app only |
Sources
OpenAI, Meet dots and Control your dot (learn.chatgpt.com/docs/dots).
OpenAI Help, Getting started with your dot; Dots privacy, security, and safety FAQs.
Facts and decisions carried over from
ai-visual-deep-dive/plans/chatgpt-dots-deep-dive.md(re-checked 2026-10-09).
