OpenAI just dropped a 10-page PDF explaining how always-on agents actually work-and why prompts are dead:
the shift: your chatbot waits for you. this agent doesn't
the formula: Always-On Agent = Model + Computer + Memory + Rules
here's the 6-layer architecture:
layer 1 → its own computer - keeps working after you close the laptop
layer 2 → reach - Slack, Teams, voice, 4,000+ apps. it lives where your work lives
layer 3 → memory - correct it once, never explain it again
layer 4 → idle research - reads everything while you sleep, touches nothing
layer 5 → rules - act, ask or hand off. you decide before it runs
layer 6 → review - nothing risky ships unchecked. stop is not undo
the result: 0 of 16,600 attacks got through. but one agent with 10 jobs drifts 2x more. one agent, one job
this 10-page PDF is the blueprint for agents that work without being asked
bookmark it, then read the article on giving your Dot real autonomy ↓