Build a Meta Ads AI Agent: The System Running My $1M+ Ad Spend

Key Takeaways

If you want to run your ad accounts with a system that works while you sleep, I'm going to show you exactly how we built meta ads agents using cloud code. Because the truth is, if AI has left you underwhelming, it's not the model. You're just using it the wrong way. In this past year, I bet almost everything on one idea. stop prompting and start hiring AI. And today I'm taking you under the hood, the real architecture, the actual code, and the one thing you can steal and start implementing this week. So, let's hop into it. And by the way, this exact system I use to run real money, one million, a little bit more than a million dollars a month these days, it helps us manage all of that ad spend on Meta alone across our ecom clients with a tiny human team. So, this isn't theory anymore. I want to show you how I built it. how these agents are running everything from big cap utilization to creative strategy handoffs, agents communicating together. I'm going to show you it all. So, without further ado, let's hop right in. But be first, we have to fix this one thing almost everyone gets wrong. If you've watched anyone talk about AI recently, you've heard the same thing. Open a chat bot, paste a giant prompt, hope you get it right, then you're like, "No, this isn't what I was looking for." And the AI is like, "Yes, it is. Oh, you're right. I'm sorry. Here's another version." Like, man, this still isn't correct. Because the problem is that a chatbot, how many people use GPT or Claude or Gemini, it's not a system. Every variable is going to be different almost every time. A prompt has no memory, no rules, no procedures. It's a brilliant intern with amnesia or context window problems. Every time it's going to guess from scratch. So, here's the shift that changed everything for my agency. We didn't build a chatbot. We hired agents, a team of them. Every agent is built like an employee, not a command. And an employee has four things, four layers, a soul, knowledge, SOPs and process documents, and guard rails. And that's the whole framework, guys. Like, if you take nothing else from this video, screenshot this part right now. And I want to walk you through how each one the real code behind it and how we use it. Because without these your agents are freestyling and you're still using chatbt or claude like a chatbot. So there's layers. Layer one, most important is the soul. This gets injected into the prompts. So the AI or the LLM model that you're using automatically knows, oh, I'm the Facebook ads agent. I have to look for these things. These are what I care about. Ours aren't blank assistants. The meta media buyer has a name, a role, a personality, and we call it Rockefeller because it manages the money. It leads with numbers, doesn't sugarcoat bad campaign. It flags a problem before it burns the budget. You're talking to a specialist with a point of view. And it's not a metaphor. It is real life. Open. Look at this file, right? We have Rockefeller, meta ad strategist, version 2.1, token budget, identity, personality, expertise. All of this is the persona, and it's the cheapest way to layer copy. So, there's no excuse not to give your agent a soul. Layer two is the knowledge, what it actually knows. This is where we wrote down years of my SOPs from media buying millions and millions of dollars on Meta. This isn't the blog version. This is literally the real version. How the algorithm behaves, how creative fatigues, what time frame windows to look for, attribution, if an ad has late attribution to know to turn it back on. The intern ha now has a decade of expertise of mine loaded directly into it before it ever answers a question or runs one of its automated jobs. Layer three is where almost everyone misses SOPs. Guys, as I've built these agents, I truly understand that training an AI agent is the same way you need to train a team member because without the SOPs, they have no idea what to do. the SOPs, bid caps, cost caps, target rorowaz, time frames to look at, three day, 7day, 14-day baselines, yesterday's performance. All of this stuff is linked directly to the SOPs. So, anyone can rent the same model. We all can use Opus 4.8 that dropped last week. but nobody else has our procedures, our documents, how we've achieved these results for our clients. And so this is one screenshot of one of our SOPs, scale provering winners inside CBOS. CBO is X% of our account media plan. I'm not giving you away all the sauce because you're not going to see how we look at these different metrics, so I've blurred them out. But this is exactly the framework that we're using. So this is just one example of the SOP baked into our Facebook ads agent. So CBO scale campaign needs to have x% of the media plans budget audiences. All of these things exclude these purchasers kill based on this SOP and kill window. Layer four is going to be the guard rails. And this is super important because we've seen people with crazy API bills and we've seen people who are like, "My system went completely crazy or deleted my entire database." We've all seen these. This document in the MD file is going to be super important. These are things it must never do. Every agent, no matter how confident I am, obeys a shared set of rules that it can't break. For example, it can't increase budget more than 20% without my authorization via Slack approve button. It can't pause more than X amount of ads within a single adset without my button approval in Slack. The brakes are exactly what has let us drive this on real money. And now here's where actually makes it reliable. These four layers don't get loaded in once and forgotten. They get assembled into every decision. Not just the first one, but every single one of them. It holds the same standard at 7 a.m. Monday as Friday night. The consistency is something that a human team literally cannot match. Like that's the thing with these agents is we start to have the ability to follow direction more than a human ever would. So let me show you the actual line that it does. These are the soul loaders. So this is where it is extremely important. Every decision is composed of four layers. All of these the interface soul identity who we are their limits their expertise the persona these get injected fresh on every single decision that's the whole system into one line it's not magic it's literally just cloud code building these things knowing that through troubleshooting and all of this and shadow gates all of this different code is how we now have really effective meta ads media buyers or agents and they share a brain. So when we learn something new, a tactic, platform change, a mistake we don't want to repeat again, we don't want to retrain 10 different bots. We write it down once and every agent that points at it gets smarter at the same moment. So we teach once and the whole team learns. Now because we've built this infrastructure on our own, we didn't take open claw or Hermes. We've been able to take different pieces of these various repos and install them into our system because they're MIT licensed. So, Paperclip, for example, was a perfect way of watching how the orchestrator layer works. Orchestrator agent calls the sub agent who then has the ability to talk to this agent, but they're kind of gated over here. Paperclip was a perfect example of using an org chart to define our agents. Reporting agent, media buying agent, Google agent, creative strategist agent, graphic design agent, copywriting agent, email agent, and communication agent. All of these are working because the certain agents point at the same knowledge module. So, we update that one line and every agent referencing it levels up immediately. Unlike every tool you've bought, this isn't frozen, right? Like this is a self-learning loop. It compounds. And remember, this was never a one genius agent. It is a roster like I just talked about. Meta, Google, creative reviewer, copywriter, graphic designer, reporting on top of them sits a single orchestrator that is the manager agent or the COO agent handling handing off the correct tasks to the right specialist. So, you're not babysitting one tool. I'm literally have a department now that runs the agency. Now, the rule I want buried into your brain before you point this at a budget, the AI proposes and the human, me or a team member, still confirms. Nothing executes on its own. Not yet. I might open up to like allow for pauses based on CPA and kill thresholds, but for now, it's one click from the human. We are still the thinkers. Does this make sense? has this ad really fatigued? All of these different things that the agents are now doing based on very strict rules. We still control whether the action gets done or not, which is why it allow I allow this system to run millions of dollars a month in ad spend. So, enough slides. I want to show you guys behind the scenes a little bit on this these agents and how they work. Now let me start because you can do a version of this today. No stack required. Open one document. Make four headings soul knowledge SOPs guard rails. That is it. That is the skeleton of a working agent. Everything we have built has just become more powerful version of these four different MD files. That is critical. So start with the soul. Name it. Give it a role. Point of view and one thing it cares about more than anything. Five lines is plenty to start. But if you want more info, you can join my school community. Link is down below. The second your an AI knows who it is, the quality of every answer will start to improve. Now, the SOPs, this is the step everyone skips, but do not skip it. Please do not. If you don't know how to run Facebook ads, watch a bunch of YouTube videos, get some of the transcripts, put them together, ask for processes on how to run ads, and then convert that into an SOP folder within your agent hierarchy. Then it will read the SOPs. That's how things become something that AI can run on. Then your guard rails. Just three to five hard rules. Never blow the budget without human approval. Never pause a proven winner. Don't touch a campaign still in learning. The one rule above them all the AI proposes and you confirm. Those are the breaks that let you move fast without blowing yourself up or spending way too much money on Facebook ads. And the last step, this is the fun part. Paste the doc into your system prompt, your project instructions, custom GPT, whatever you're doing, wherever the model reads. And here's the magic. Every time it makes a mistake, don't retrain it. You just add one line. Do that for 90 days and you have a document that your agents will use that nobody can copy because it's built from your money, your repetition, your managing the ad spend. So, here's what I want you to take home. Stop writing one giant prompt. Give your agent four things. A soul so it knows who it is. Knowledge so it knows your world in which we live in. SOPs so it can do things your way and guard rails so it doesn't go off the rails. You don't need a full stack to start. Even a simple version written in a doc, an MD file beats the mega prompt you've been writing and still not getting right in chat GPT or clot. So that's how you can scale in 2026. Agents are here now. The key is to start building them so they can get better. You can start by more creative, sharper decisions in front of the algorithm. And if you want this to hit every time, there's something underneath it all you're going to understand. Most operators are still running ads like it's 2024. Watch in the next videos. I'm going to break down the Andromeda update, but also if you want more information on this, if you want to see how I've built our entire stack, check out my school community down below. Seven days free. Let me know if you came from YouTube and I'll get on a 30 minute call with you to really break down exactly how to get this started. We'll see you guys in the next video.