korrents

korrents · video

Make Your Agent Safe and Autonomous

Kent C. Dodds · 24m · youtube.com ↗

9 korrents from this recording

Kent C. Dodds did not write this page.

Every claim below is a statement made in this recording, quoted word for word and linked to the second it was said, so you can hear it rather than take our word for it. The wording comes from the transcript published alongside the recording; the sentence above each quote is our reading of the claim, not their wording.

  1. 0:06 · watch on youtube.com↗

    Uh, actually having AI write a first draft of an email is pretty nice, but having it send that email without you reviewing it can be irritating.
  2. 1 min later
  3. 1:21 · watch on youtube.com↗

    If you watched my video the while back, you saw how an agent was able to drop a production database because it was an overscoped token.
  4. 5 min later
  5. 6:13 · watch on youtube.com↗

    And this is a a general practice that I would recommend is that you can just ask the agent if you ever get stuck on anything. Just tell the agent that you're stuck and ask it what to do next.
  6. 6:38 · watch on youtube.com↗

    and the more and more capable you make your agent, the more it can do kind of on its own and you don't have to control every piece of what it's doing and and be the like context hot potato person.
  7. 5 min later
  8. 11:08 · watch on youtube.com↗

    When you run into a problem with your agent, you can literally just take a screenshot here. We'll we'll do the full thing right here. Paste it into your agent without any context at all. Just here's the screenshot and it will figure out what's going on and tell you what you should do next.
  9. 12 min later
  10. 23:14 · watch on youtube.com↗

    Yes, we trust the agents to act in good faith and stuff, but sometimes they misunderstand us. And if they have the power to drop your production database, that has happened and it could happen.
  11. 23:24 · watch on youtube.com↗

    So you want to make a package of deterministic code that can only do the things that you know you want them to be able to do. And then you can lock down the integration and lock down the package so that agents aren't able to do more than the things that they should be allowed to do.
  12. 23:52 · watch on youtube.com↗

    I want you to take a look at your&env file and look at all those tokens that your agent is happily slurping up and able to use and uh figure out how to get those integrations into a safer place where agents won't be able to just willy-nilly use tokens uh however they want to
  13. 24:19 · watch on youtube.com↗

    And uh I just want you to be a little bit more intentional about what your agents are able to do so that you can reduce the chances for catastrophe.