Grok Bots: Stanford Lecture · Open Campus: AI for Business

Grok Bots: Stanford Lecture

A terminal harness dies when you close the laptop. This lecture is the next step: an agent that keeps working after you leave.

Lee Robinson (ex-Vercel, ex-Cursor, now SpaceX) walked Stanford CS146S through how Grok Bot is built. About 45 minutes. The useful part is not the product. It is the harness pattern underneath it.

Always-on does not mean the model thinks 24/7. The server and the cloud computer sleep. A message, a communication ping, an email, a meeting, or a routine wakes it, loads state, runs one turn, saves, and goes back to sleep.

Four eras, in his framing:

Copy-paste out of chat into an editor
A model that can edit files and run shell in a loop
Many agents in parallel, with you still approving every command
Always-on agents that run in the background, with their own computer and memory

His close is the part that matches what we are building here. Assume today's models are the worst they will ever be. Build as little product as you can get away with. Write the process down so an agent can run it. The new primitives (durable workflows, sandboxes, memory, skills, evals) are the stack, not another chat box.

If you only have 10 minutes: start at the harness section, around 18:00. The context-engineering section at 27:00 is the one that will change how you run your own agents.

What are you still approving by hand that a quiet background agent should already own?

0 comments

Grok Bots: Stanford Lecture