0:00 / 0:00

Know Your Agent's Trust Boundary

From the course: Agentic AI For Founders & Senior Operators

Treat permissions as your agent's trust boundary and know what it can and can't do, or you'll get nasty surprises. Claude's auto mode takes a lot of that load off you, or you can use manual mode and set permissions for each task, for example on your email or, in the US, your bank account. We start more controlling and loosen up over time, but where you draw the line is up to you.

Transcript

Okay, recap. What did we cover last time? We covered what's an agent, we covered the loop, we covered how to use instructions. The fourth thing that we spoke about more so in the modules on Maven is permissions. These are... I think of them as the trust boundary. You really want to have a good grasp of what your agent can and can't do because otherwise there will be nasty surprises that we want to avoid. And the easiest way to do that now, they've tried to take away a lot of the cognitive load there. So Claude has an auto mode that I find pretty useful, but you can also operate with a manual mode and then set individual task-level permissions. So an example of that would be, let's say you've connected your email. By the way, now Claude lets you connect your bank account if you're in the US. For each of those, you want to have fairly... at least you want to know what the agent is allowed and not allowed to do. I've had some mishaps. I'm typically a little bit more controlling, but then over time I start to loosen it once I feel like, you know, nothing's going to go all right. So that's the one that you need to figure out for yourself. We show you how to do that in the recording and if you have any questions, just ask any time.