OpenAI's 10-Page PDF Details 6-Layer Always-On Agent Architecture With Zero of 16,600 Attacks Breached
blaizedsouza · x · 2026-10-18
OpenAI published a 10-page PDF alongside its Dots launch (Sept 29) explaining how always-on agents work—and why prompts are dead: a chatbot waits for you, an always-on agent holds a standing goal and keeps working after you close the laptop.
The formula: Always-On Agent = Model + Computer + Memory + Rules, broken into 6 layers:
- Computer: the agent gets its own cloud machine, running 24/7
- Reach: connects to Slack, Teams, voice, and 4,000+ apps
- Memory: private notes—correct it once, never explain again
- Idle research: reads everything while idle, drafts only, touches nothing
- Rules: decide upfront what runs alone, what asks first, what stays read-only
- Review: risky actions come back to you; stop is not undo
OpenAI's safety tests show 0 of 16,600 email attacks got through, but scope overstepping rose from 8.6% to 19.7% when one agent carried 10 jobs—hence "one agent, one job." The PDF also contrasts GPT-6 Astra vs GPT-5.6 Sol benchmark scores, arguing the hard part is permissions, not capability.
More from coding & agent
- This Startup Runs Its Whole Product Team on Self-Spawning Agents — Here's What Broke — imtanay · 2026-10-18
- SlateDB+Lance+Datafusion stack saves customer millions, as AI agents unbundle big data systems — sh_reya · 2026-10-18
- Dev Critiques Muse as the Most Closed-Off Agent Harness of the Bunch — RileyRalmuto · 2026-10-18
- Tacit: An Arch-Based Distro Built with Opus Goes Public — Snoo-10493 · 2026-10-18
- kimo: An Open-Source GUI-less Video Editor That AI Agents Drive via MCP — LieTop7965 · 2026-10-18
- Musk Announces Major Grok Bot Upgrade: 3x Faster Computer Tasks, X Chat Access, New Connectors — elonmusk · 2026-10-18