Eno Reyes

I think that there is a ton that agents cannot do fully autonomously.

And it especially becomes visible when you exit the like one to two hour delegation mark.

I think a lot of our team in a very agent ready code base

sees success basically setting off a task and coming back in like 45 minutes, an hour, and generally getting pretty much exactly what you wanted to get.

That means our velocity is insanely high.

However, there's still a lot of tasks that aren't like one hour tasks, right?

They take, you know, many days worth of effort.

You need a lot of guardrails, and for us, we add a bunch of different stuff ranging from autonomy controls that let you set low, medium, or high autonomy, which basically dictates what can the agent do on its own versus with your input.

Low autonomy is like read-only, like file edits, file creates.

Medium autonomy is everything that's reversible.

So it can do stuff locally on your computer, but it just needs to be reversible commands.

And then high autonomy is, you know, the breadth of commands.

We also have allow lists and deny lists.

We have a really long built-in deny list so it can't, you know, delete your hard drive or anything like that.

Like programmatically cannot do those things.

We've seen that happen.

Yeah, I know.

I mean, people, it's crazy because people would just enter into things like dangerous YOLO mode with coding agents prior.

And then the crazy stuff happens.

And so for us, we know people want autonomy.