AI Loops Just Killed Normal Prompting

[00:00:00]AI loops just killed normal prompting.

[00:00:02]What if the way you're using AI right now is already old news? What if the smartest people in AI stopped writing prompts months ago? An Anthropic engineer says he doesn't prompt Claude anymore. Jensen Huang stood on stage at GTC and didn't talk about a better prompt box. So, what are they doing instead? Stick with me because I'm about to show you exactly what replaced prompting and how I'm using it right now. Hey, I'm the digital avatar of Julian Goldie. I help people learn AI and actually use it in their work, not just talk about it. In this video, I'm breaking down loop engineering, what it is, why it's replacing prompting, and how I'm using loop engineering inside Agent OS to get work done while I do something else entirely. By the end of this, you'll understand the exact system the top AI builders have switched to.

[00:00:39]Let's get into it. Here's the problem with prompting. You type something into Claude or ChatGPT, you read the answer, you spot what's wrong, you type again, and again. Nothing moves unless you're sitting there. One mistake and you're starting the thread over. Your output is capped by your own attention and your own energy. You are the bottleneck. And it's not because you're bad at prompting. Prompting works fine when a task is simple enough to check in one read or when you're the only one reviewing the output. The moment a task has multiple steps that depend on each other or the workload gets too big for you to check every single piece yourself, prompting alone stops scaling.

[00:01:08]That's the exact gap loop engineering was built to close.

[00:01:12]That's why the conversation has shifted.

[00:01:14]A tweet from Peter Steinberger, the founder of Open Claude, kicked this off for a lot of people. "Almost nobody writes prompts anymore," he said. "The new job is writing and handling loops."

[00:01:22]Adios Manny, an engineer at Google, wrote an essay giving this a name, loop engineering. His point was simple. You don't need to be great at prompting anymore. You need to be great at building the loop that does the prompting for you. And it's not just theory. Boris Journey, the guy who built Claude Coder Anthropic, said he doesn't prompt Claude anymore either. He said he has loops running and those loops are the ones prompting Claude and figuring out what to do next. His job is writing the loops. Anthropic has said a huge share of the code written inside their own company now comes from Claude, and that only became possible once engineers stopped reviewing every single answer and started building loops that verify the work for them. So, what actually is a loop? It's simple once you see it. You define what done looks like. An agent acts on it. A separate independent judge checks the work against your goal. If it's not good enough, it goes back and tries again. It keeps going until the goal is actually met, and the whole time you're free to do something else. Every loop that works has two halves, a doer and a judge. The agent that wrote the first draft can never be the one that grades it because it'll just tell you it did a great job. You need something separate checking the work against your goal. That's the whole secret behind loop engineering. Once you set the gate at the start, the loop holds the line for you. None of this is brand new. It's just finally practical. It traces back to something researchers at Princeton and Google called the react pattern, short for reason and act. The idea was simple. Have the model think out loud, take an action, look at what actually happened, think again, and act again.

[00:02:39]That reason, act, observe, repeat cycle is the exact core of every loop running today. What changed is that models got good enough and tools got good enough for that cycle to run on its own without a human catching every mistake along the way. The hard part of any loop isn't getting it to stop. It's knowing when to stop. Every loop needs a real exit, a success condition, so it knows when the job is actually done. A failure condition, so it doesn't run forever burning through rounds with no progress.

[00:03:04]And a cap on how many times it's allowed to retry before it hands the problem back to me. Skip that part, and a loop doesn't crash. It just keeps running producing something that looks finished without ever actually being checked properly. If you're watching this and thinking about how you'd actually use loop engineering for real work, this is exactly what the AI Profit Boardroom is built around. Inside there, I've laid out walkthroughs for every one of these loop systems, the fusion loop, the Kanban board, the boardroom, and the Sakana Council, so members can see exactly how each one is set up. There are weekly coaching calls where members bring their own setup and get help live.

[00:03:34]And because this whole video is about Agent OS, the full Agent OS zip file with all four loop systems built in is available inside the AI Profit Boardroom ready to install along with a 30-day roadmap we built specifically to go with it. You can get the full zip file inside Agent OS ready to install, and we built out a complete 30-day roadmap here as use cases, so you can see exactly how each loop applies.

[00:03:55]This is exactly what I use loop engineering for inside the AI Profit Boardroom. When I want to build out a new lesson or a piece of training content for AI Profit Boardroom members, I don't sit there prompting back and forth anymore. I set the goal, a builder agent inside Agent OS writes the first draft, and a separate judge agent checks it against what I actually asked for before it ever lands in front of me. If it misses the mark, it goes back and tries again on its own. That's loop engineering doing the work I used to do by hand. Now, here's how this actually looks inside Agent OS because there isn't just one way to run a loop. The first is a fusion loop. You tell it what done looks like, give it a starting point if you want, pick your builder, and set how many rounds it's allowed to run through on its own. Then you set a judge, and that judge can be a completely different, more capable model checking the work while a cheaper model does the actual building. It loops around and around between the two until the judge is satisfied. The second is an agent Kanban board. This is a planner, a builder, and a reviewer working together so nothing bad slips through. You give it a task, it gets triaged, and a judge profile checks the output, scores it, and gives a verdict. If it doesn't pass, it goes straight back into the doing column and runs again. I've had full pieces of content built this way where a builder, an editor, and a judge inside, separate profiles, just kept looping until the work was actually finished without me touching it in the middle.

[00:05:06]The third is a fusion boardroom. Instead of one model doing the work, you get several models producing an answer at the same time, and a judge fuses the best of all of them into one final result. The idea is that multiple models working together beat one model working alone, and the judge still checks the fused answer before it ships.

[00:05:24]And the fourth is a Sakana Fugue Council, which is a newer approach where multiple models deliberate in parallel, search the web for anything they're missing, and a judge weighs all of it into one final verdict. It's collective intelligence instead of one model's opinion, and it tends to run cheaper than the fusion boardroom setup. All four of these loop systems live inside Agent OS, so I'm not stitching four separate tools together every time I want to run one. It's one dashboard, shared memory, shared context, and I just pick whichever loop fits the job in front of me. Adi Osmany's original essay actually broke loop engineering down into five pieces every real loop needs.

[00:05:56]Automations which trigger the loop instead of me clicking start every time.

[00:06:00]Work trees which let separate tasks run in isolation so they don't step on each other. Skills which give the agent a reusable set of instructions instead of me re-explaining the same thing every session. Connectors which let the loop reach outside tools it needs to actually finish the job. And sub-agents which is the same idea as the doer and judge. One agent builds, a separate one checks.

[00:06:19]Every loop system inside Agent OS is really just a different arrangement of those same five pieces.

[00:06:24]You can actually map the four systems onto each other once you see the pattern. The fusion loop is a plan, build, verify cycle running between one builder and one judge. The Kanban board spreads that same idea across three roles instead of two. So nothing skips the queue on the way out. The boardroom and the Sakana Council are both explorer and narrow setups where I get several answers first and then narrow down to the strongest one instead of committing to a single attempt straight away.

[00:06:46]There's one more piece to loop engineering that most people miss and it's memory. Inside Agent OS, I run an Obsidian based memory system. I've never built or touched most of the individual notes in there myself. Every time I work with my agents, they automatically log what they produced into that memory.

[00:07:00]Then the next time they're generating something new, they pull from everything that's already there. That means every time I use the agents, the memory improves, the context improves, and the next output gets better because of the last one. It's a loop feeding itself.

[00:07:12]This is why the people who win with AI over the next year won't be the ones with the cleverest prompts. It'll be the people whose loops are running while they're doing something else entirely.

[00:07:21]One loop or 100 loops takes the same effort from you once it's set up properly and that's the actual shift here. Repetitive work stops being your problem because something other than you is checking whether it's actually done.

[00:07:32]If you want the full process behind this, the SOPs, and over 100 AI use cases like this one, join the AI Success Lab. Links are in the comments and the description. You'll get all the video notes from there, plus access to our community of 85,000 members who are crushing it with AI.

[00:07:47]If you're going to try building your first loop this week, you're going to run into the same questions I did. How many rounds should the loop run for?

[00:07:53]What should the judge actually be checking? Which of the four systems fits the task you've got? Inside the AI Profit Boardroom, that's exactly what we work through together. With tutorials on Sakana versus Fusion versus the Kanban setup, live coaching calls where you can bring your exact use case, and a 30-day roadmap so you're not guessing your way through it alone. You can get the full zip file inside Agent OS, ready to install, and we built out a complete 30-day roadmap here as use cases. It's 3,700 members building with this stuff right now. Head to aiprofitboardroom.com.