korrents
Lee Robinson

What Lee Robinson thinks about AI agents

@lee-robinson · 58 positions · On ppll ↗

Works on ML at SpaceX, as of 2026; formerly at Cursor, and five years at Vercel before that, ending as VP and working on Next.js. Writes at leerob.com.

Lee Robinson did not write this page.

We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.

18 dated positions, 2025 to 2026, in their own words. Our reading of what Lee Robinson has said — not written or endorsed by them.

  1. a separate AI model that's checking all of the commands that get run. So, you know, kind of funnily enough, that's actually more secure than the fatigue of having to watch hundreds of shell commands and approve each one.

    ↗Stanford CS146S: How always-on agents workyoutube.com 1st of 22 in this recording

  2. MCPSaaS

    otherwise known as computer use, which basically means that any SaaS product or any website that didn't expose an API or an MCP server can be used headlessly, basically.

    ↗Stanford CS146S: How always-on agents workyoutube.com 5th of 22 in this recording

  3. Um but the most critical thing here is that if you put the agent loop on the virtual machine, then you can't shut down the virtual machine independently from the uh from the agent server.

    ↗Stanford CS146S: How always-on agents workyoutube.com 6th of 22 in this recording

  4. So, what you really want here is you want a durable workflow, which means that it's safe for you to actually retry that, and when you retry after a failure, it picks right up on step three.

    ↗Stanford CS146S: How always-on agents workyoutube.com 7th of 22 in this recording

  5. In general, the principle here really is to try the cheapest option or the most reliable option first and kind of work your way up the ladder here.

    ↗Stanford CS146S: How always-on agents workyoutube.com 8th of 22 in this recording

  6. Like, okay, this thing failed. Like, click on this link or here's what you need to do next. Turns out that's also very helpful for agents.

    ↗Stanford CS146S: How always-on agents workyoutube.com 10th of 22 in this recording

  7. eventually it's going to hit a point where it needs to summarize and all compression is lossy. So you have to figure out the best algorithm to do the compression and not lose important details

    ↗Stanford CS146S: How always-on agents workyoutube.com 13th of 22 in this recording

  8. I think it's very possible that pretty soon your agents will be working for weeks or months. And it does bring up some kind of interesting questions in terms of how we train and evaluate those models.

    ↗Stanford CS146S: How always-on agents workyoutube.com 15th of 22 in this recording

  9. I think ideally if you have this digital colleague, you give it literally all the tools that you have if you were onboarding a real human. And so I think organizations are still warming up to that idea, of course.

    ↗Stanford CS146S: How always-on agents workyoutube.com 16th of 22 in this recording

  10. what do you do when you need to evaluate an agent that runs for a month, but the models ship every month?

    ↗Stanford CS146S: How always-on agents workyoutube.com 17th of 22 in this recording

  11. uh, I don't think that we have it perfect yet and I think trying to thread the needle between being proactive and suggesting

    ↗Stanford CS146S: How always-on agents workyoutube.com 18th of 22 in this recording

  12. And the reality is, this is just kind of the new normal now where every new product is going to have to have some kind of agentic side to it, either integrating or it will become a you know, a headless piece of data for another agent to use.

    ↗Stanford CS146S: How always-on agents workyoutube.com 20th of 22 in this recording

  13. 6 months earlier
  14. And I typically recommend if you're starting a new feature, you should probably start a new conversation with the agent. Or if you see the agent kind of going off track and making mistakes, that's also a good sign to start a new conversation.

    ↗Cursor: coding agents tutorial (2026)youtube.com 2nd of 18 in this recording

  15. if you're searching through a very large code base with many files, it's going to generate a lot of context usage and a lot of tokens. So, I highly recommend using sub agents to improve your context management.

    ↗Cursor: coding agents tutorial (2026)youtube.com 4th of 18 in this recording

  16. I find that when you break it down into more semantic commits, it's much easier for a human reviewer to step through the commits and understand things if they have more questions.

    ↗Cursor: coding agents tutorial (2026)youtube.com 13th of 18 in this recording

  17. A great rules file is short, it's specific, and it points to examples in your code base that you want the agent to reference.

    ↗Cursor: coding agents tutorial (2026)youtube.com 16th of 18 in this recording

  18. Now, you might be tempted to write rules for everything, but you should really resist this. Too many rules will consume just an unnecessary amount of context, and it might honestly confuse the agent.

    ↗Cursor: coding agents tutorial (2026)youtube.com 17th of 18 in this recording

  19. 8 months earlier
  20. and the reason why I like to show this example um is that deterministic things are good. Like code mods are helpful. We know that this is going to solve some things very concretely and then if there's errors or issues, we can still use AI and use the AI agent to help us fix some of those things.

    ↗Refactoring a Next.js & Tailwind app with Cursoryoutube.com 3rd of 9 in this recording