Bryan McAnulty

It has to respond with something about that.

And so what the AI is doing in its training is it's trying its best to do whatever you said in the next minute and deliver something.

It may be what you would determine is actually half done or actually incorrect.

But if you can realize that that's what the AI is trying to do, and then after that, if you put it in the loop, it's technically another AI.

You might be feeding it the context of what happened, but in a way, it's another AI.

So it's like being brought to life over and over again of all these different agents with these different memories that you're kind of forcing into them rather than one thing that's always working.

So if you think of it that way, I think you can start to think about how are you giving it the right information so that way it can perform the test that you're looking for.

Yeah, well, it doesn't even have to be as complicated as having to technically set up some kind of sub-agent or something like that.

What it's about, again, is directing the attention.

And so if you're building the software or building some kind of thing, then you can say, okay, this is the thing we're building, and then maybe the next equivalent agent, you don't have to say anything about the agent, you're just defining it in a long prompt.

Where it's like, okay, then we need to check over all this for security, or we need to check over this for optimization, or we need to go and do some research on the web to make sure this aligns with the marketing thing that we're trying to do.

And so it's distributing where are the things that we think are important to kind of spend some attention.

I think it comes back to, again, just building that sense of actually trying these things out and working with the models, because when I saw everyone talk about looping, I saw the founder of OpenClaw says, you should be working on building these loops.

This is the future.

I would actually push back on that a little bit and say that it's not just about building loops, because if we have the infinite loops for everything, then everybody just has a slop factory, right?

And so we have to realize like where are the parts that we want to have a back and forth where we're iterating on something that we care about.

We need to see, okay, what does the interaction here look like?

Or we need to see some kind of information before we can tell the agent to continue versus something where we're able to articulate like this is a very clear thing that needs to be done.

And the minor specifics of how something needs to be accomplished is not so important.

That's the kind of thing like, okay, the agent can just be working on this in the background.