How always-on agents work

Lee Robinson (leerob) · YouTube

For anyone building agents other people rely on, Lee Robinson's Stanford lecture lays out what it takes to keep one running after the laptop closes.

What holds up

Durable workflows, not job queues: when a task crashes, it picks back up at the step that failed instead of replaying from step one.
Thin client, thick server: triggers like Slack, email, meetings, and routines wake the agent on a server, which spins up its own computer only when it needs one.
Most of the talk is context engineering: helper agents keep the main conversation small, and prompt-cache discipline is treated as part of correctness. He calls a cache miss "kind of like a bug in the system."