Stanford CS146S: How always-on agents work

Lee Robinson's Stanford CS146S talk on how always-on agents such as Grok Bot work: server-hosted agent loops with their own cloud computers, durable workflows, dynamic tool discovery, helper subagents, prompt caching, compaction, layered memory, and the training flywheel behind them.

4 sections ยท 12 lessons

Course outline

From Loop to Colleague

  1. From Copy-Paste to Always-On
  2. Six Shifts That Made Bots Possible

Always-On Architecture

  1. Sleeping Servers, Waking Triggers
  2. Thin Client, Thick Server
  3. Durable Workflows Over Job Queues
  4. Designing a Meeting-Notes Bot

Harness Engineering

  1. Try the Cheapest Option First
  2. An Orchestrator With Helpers
  3. Keeping the Prompt Cache Warm
  4. Compaction and Layered Memory
  5. Why Did the Bot Bill Triple?

Trust and What Comes Next

  1. Guardrails for Unattended Agents
  2. The Harness-Model Flywheel
  3. Delete the Product

Start learning with Wondering