Stanford CS146S lecture: Lee Rob explains how always-on AI agents work
NielsRogge · x · 2026-10-18
Lee Rob's Stanford CS146S lecture "How always-on agents work" is now free on YouTube, detailing the infrastructure behind persistent AI agents.
Key architecture points from the slides:
- Agent's computer: one Firecracker VM per user, saved to disk when idle; runs a Linux (Debian) container kept alive by a supervisor
- Components: file server (files, app connections, live desktop view), command runner (shell commands, screenshots, clicks), and multiple virtual screens (one Chrome per active agent)
- Sleep: a few minutes after last activity, never mid-turn — files sync to cloud storage, memory and disk saved as a snapshot, machine stops
- Wake: on the next tool call, snapshot restored in place; logins and installed tools persist, conversation state recovered
The core principle: files, installed tools, and browser logins all persist between sessions — the infrastructure foundation for truly always-on agents.
More from coding & agent
- Lawyer builds agentic interview tool that maps legal workflows for AI pilots — jkubicki · 2026-10-18
- Smash dev recants anti-AI stance after learning 90% of coding is now AI-assisted — IanArawjo · 2026-10-18
- From company brains to company harnesses: a shift in enterprise AI thinking — jacob_posel · 2026-10-18
- Dev builds offline Arch distro with Codex to revive orphaned packages for AI — Maleficent_Platypus6 · 2026-10-18
- Curlity: A Visual HTTP Viewer MCP So You Can Audit Your AI Coding Agent's API Calls — RefactoringReality · 2026-10-18
- Reddit thread crowdsources the real failure modes of frontier coding models like Opus 5.5 and GPT-6 — SadPossibility2338 · 2026-10-18