Stanford CS146S: How always-on agents work

Name: Stanford CS146S: How always-on agents work
Uploaded: 2026-10-16T00:00:00.000Z
Duration: 2717 s
Channel: leerob
Description: Lee Robinson's Stanford lecture walks an always-on agent's architecture: a server-side loop on Firecracker microVMs, Temporal workflows, a one-tool thin client, prompt-cache discipline, warm-window compaction, three-layer memory, autoreview, building for the next model.

Lleerob

Full transcript

English

Summary：Lee Robinson's Stanford lecture walks an always-on agent's architecture: a server-side loop on Firecracker microVMs, Temporal workflows, a one-tool thin client, prompt-cache discipline, warm-window compaction, three-layer memory, autoreview, building for the next model.

Watch on YouTube

Core points (3)

1The thin client exposes exactly one tool; any new surface attaches at zero cost.
2Prompt-cache discipline is a first-class constraint; a cache miss is treated as a bug.
3Delete the product: build for the next model, assuming today's is the worst.