Stanford lecture breaks down Grok Bot's always-on agent architecture
A former Cursor engineer explains in a Stanford CS146S lecture how always-on agents like Grok Bot work: always-on cloud infrastructure, event-driven execution, persistent memory, sub-agents and context engineering.
This is fuc*ing ridiculous The next AI breakthrough isn't a smarter chatbot. It's an agent that never clocks out. @leerob, an ML engineer at SpaceX and former Cursor engineer, just shared a Stanford CS146S lecture explaining how always-on agents like Grok Bot actually work. The key insight: the model isn't the whole agent. The infrastructure around it is what makes it persistent. Here's what changes: → Always-on infrastructure: A dedicated cloud computer keeps the agent ready to act in the background. → Event-driven execution: The agent wakes when something happens, completes a turn, then stops until the next trigger. → Persistent memory: State lives in a database, so crashes don't erase progress. → Sub-agents: Heavy tasks get delegated to fresh contexts, keeping the main agent focused. → Context engineering: Profiles, dated logs, and reusable skills give the agent the right information at the right time. → Selective communication: A dedicated send tool lets the agent work quietly and report back only when needed. The smartest design choice might be what the agent doesn't do. It checks memory before searching, uses APIs before resorting to browser automation, and delegates work instead of bloating its context. The future isn't AI that answers every prompt. It's AI that manages ongoing responsibilities. Watch the 45-minute lecture below to understand the architecture behind agents that behave more like digital colleagues than chat windows. https://x.com/leerob/status/2108650243365736855/video/1