This pattern is crazy, or I'm an idiot - Learning Domain-Driven Design by Vlad K [58:06-63:06]

or does not function that way, right? It's very test driven development sounding. But for me, I've been spending a lot more time. What I've done in my best jobs in the past, but doing this now is building little command line tools because there's a lot of stuff that I don't want LLMs to have flexibility on. And once I kind of figure out what it flow is, I'll build a tool that gives it an instant constraint. Um, and then all of a sudden, I can trust the tool more, right? So I can be like, okay, I figured out that I always want these five things to operate any time I'm making a decision or maybe I have a, a sub agent that is a defender of everything that's in our ADRs or, or, or, or, our respective driven development. And what it does is it's in its context window, it like slurps all that stuff up and then compares what code changes we've made. Does it meet alignment with, or is there a contradiction, right? I think these kind of little games that you can play can be really useful for looking for blind spots that I have. Maybe I've forgotten like, oh yeah, we agreed that we would never structure user data this way because of this ADR that we have and Claude obviously was missing the context there and is trying to push this thing through. And I forgot about it myself. And I have the other sub agent that's kind of like sticking up for the docs, right? Or whatever, right? Those, I think, but again, you get back to these like Unix philosophy where, um, I think it was, it's Matt, uh, Matt Pocock that has brought this up really well. He tries to keep his context windows as small as possible. And so he like, he knows that the context windows are best when they're kind of, you know, three quarters empty. Uh, and so he's like, well, can I play a game in which I know that this is only pulling in context. It's really loses its full memory every time. So can I play this game where I've got this excellent paper trail. I have these very clear instructions on how to bootstrap yourself. And I have a very clear question of like the next thing I want to get done. And um, the more I've been playing that game, the more I'm like, yes, this is this sort of advanced beginner, you know, junior dev idea that we've been bringing up is that that is a great use of these tools, right? Is that you, you can go in and, um, give it some unique perspective on a bigger state of things that I'm keeping track of. And it can do those, like sort of more focused jobs really well. If I ask it to do my job, it starts to do a bad job. Right, right. We've been having not disagreements at work, but just a little like we, we actually just, we, we want the idea of like, especially for stuff that like, like low priority work, we want agents to be able to just kind of autonomously work on this stuff, which I think is really, really good. Um, and we looked at kind of like a lot like the managed solutions and didn't like any of them. And so we just had some old laptops and we're just like, just just set a cloud running on a loop on this laptop, right? And leave it plugged in all the time. Like, let's see, let's see what that does, right? Um, and so, and we've got it hooked up to linear. So as tickets come in that have a special label, it'll pick up those tickets and start working, which is good, but there's some kind of work, like, for example, with some of these queries that we have to rewrite because they, it turns out they're, they're really inefficient. And I had an engineer who was like, oh, great, like I'll, I'll, we call the laptops, the ponies. Like I'll have the ponies work on this. I was just kind of like, the ponies are going to have to like first, the ponies can't connect to the database and run the explain query. So like, we're going to have to run that for them, give it to them. And then when they run the code, or then when they rewrite it, we're going to have to pull it down and we're going to have to test it. And we're going to have to run the explain again to make sure that it works, right? And I'm just like, at this point, like, like, we're, we're farming out so little of this. And like I kind of hate this idea that like the machines will generate all the code. And then I will pull it down and test it. Like I don't, I don't have that expectation with any other developer. Like if any other developer was submitting code, it was like, I didn't test this. Can you pull it down and test it for me? I'd be like, what are you talking about? Like test your own stuff. And so that gets to do a whole other thing. We're like, I think what we need, we really need our feature branch previews, especially because like a lot of what they handle is like UI improvements. And like, that's really easy to confirm despite clicking in a feature branch. But to do that, I think we need to be on Kubernetes instead of ECS. And it's all a big thing. Anyhow, but yeah, just like, I don't know. I like they do these things to be autonomous. They need to be like actually autonomous like a junior. Yeah. I'm not quite there yet. I have some pretty cool stuff that I'm I'm working on internally. The CTO at this company that I'm working with has some really neat tools that he's been working on. And I think he's going to release it to the public at some point. It's this idea that you can mix it really simple to spin up agentic workflows