Claude Loops: Build Agents That Run Without You

**Julian Goldie** (0:00)
Today, I'm going to show you five different Claude Agent Loops that you can build to basically have Claude running without you autonomously to build and automate anything you want. Now, if you're wondering, okay, what are loops? Essentially, this means that you don't need to prompt anymore because you can give loops to your AI agents like we've got right here, and I'll explain exactly how it works in a second. And each one acts, checks his own work with an independent judge and repeats until it's actually done. Now, inside my agentic operating system, I've actually built out five different loops, one to build anything to a spec, one that works with a Kanban board, one that takes an idea from inbox to shipped, one that watches the news and one that writes your SEO, right? All running on one shared memory that makes them smarter every time. So I'm going to explain exactly how each loop works. Now, the difference here is like you are no longer the person quality controlling or checking anything, which saves a lot of time. And also these loops can run without you autonomously, or you can actually add a human in the loop, which I've also built out. So you can have two different stars of loops depending on what you want to build. If you saw the tweet from OpenClaw founder, Peter Steinberger recently, he was talking about this as well. And essentially a system where you're not going back and forth with AI anymore. This saves so much time, it just runs in the background and you can get way more done with these powerful systems. In fact, your only job now is to make sure you're the architect, right? So before, and you probably can relate to this for a year, I was the loop. So I'd type a prompt, I'd wait, I'd read the answer, spot what was wrong, type again and again, and nothing would move until I was sitting there driving it, right? So my output was pretty much capped at how long I could stare at a screen. And for a lot of people watching this, you can probably relate to that. Now what we have is a system where we build the loops ourselves, right? So I define what's done and what that looks like. An agent acts, a separate judge checks it against the goal and it repeats until its own work passes the quality control checks. Now we have five of these running inside the agent OS, building, planning, shipping, watching and writing. And they'll run whilst asleep and then I wake up to the work done. And you can run the same.
We've actually got loads of people using the agent OS inside the AI Profit Boardroom. We have over 196 pages of testimonials and wins from people using systems like this. So if I can do it and I'm non-technical, you might be non-technical too. You know, this is a pretty powerful system. So how does it work?
Well, if you look at the old way versus the new way, one makes you the engine, the other makes you the architect. So for example, with the old way, with the prompting using the old way, you type, you read, you wait, you correct on every turn. So most people are going into ChatGPT, for example, like this, and they're just going back and forth, and it takes a lot of time, it's very inefficient.
And the problem with that as well is like nothing moves unless you're at the keyboard, there's no check on the work but your own eyes, and the output is capped by your attention. With the new loop systems, you write the loop, it runs itself, you define done once, the agent acts, a judge checks it, it repeats on its own until it passes, and it runs whilst he sleeps, you wake up to it done. And every loop is the same shape, which is act, observe, judge and repeat. So it loops around and get stuff done without you. Or you can have like just a quick check to say, okay, this is good to go. So let's talk about the first loop here, which is loop engineering. So for example, if we have a look at this system, we have loop engineering. And what we can essentially do here is we define what's done.
We give it a starting point if we have one. And then we can select the builder, aka the API for the builder, how many rounds it goes round, and then which API we use for the judge. So you could have like a GLAM 5.2, which is a powerful model for the judge. And then you can have a free API for the builder. And that means, number one, you use less tokens, and number two, it costs less. And number three, this just runs round in a loop. So you can say, okay, build out a website. And we've built out several things here, as you can see. You can say, build out this project, and it will loop round however many rounds you have. The builder will create it, and the judge will judge it, and then it just loops round until the work is done. And I don't need to touch anything. And also, we can just see the outputs we've created over here inside our workspace. So it will get saved and I can check it up later.