Transcription
I just read a study that by 2030, AI is going to create 170 million new jobs, but they won't be jobs where you just sit there and chat with AI. They'll be jobs where you build AI agents.
And I get it, the AI space is moving crazy fast. I mean, what even is an AI agent? Not too long ago, I was right there with you. But after going deep myself and building dozens of agents, I found out it's actually way easier to build and manage these agents than it looks. So much so that my whole team and I have hundreds of AI agents doing 92% of all the work across my companies.
So today, we're going to go through every step on how you can build your first AI agent, starting with AI chatbot versus AI agent. A chat is like a meeting. An agent is like an employee. Chat is you ask it a question and then you get an answer. And a lot of people just copy and paste things and do something with it. With an agent, you actually tell it what you want to do and it runs the full workflow.
Think of it like these are the body parts. I call it data. So one is D, it can diagnose. It can actually figure out what the problem is and solve it on your behalf, kind of like hiring a consultant. Next is A, it can assemble. It can build a plan, it can design tools. In that way, I think of it like an architect. It knows all the different pieces that it can pull together to get something done. Next, we have T, it can take action. In that way, I think about it like somebody that executes tasks. And finally, A, it can assess. It can check its own work, see where the opportunities are, and then make sure that it landed on the right answer. And if not, it can review itself and make itself better.
This whole thing is called a loop. And without a loop, an agent would just do the job and then stop. That's called an automation. But with an agent, it keeps learning. It keeps getting better. It kind of acts like a person. With chat, it pulls on us. It's asking us, "What do you want me to do?" We prompt it and then we wait. With an agent, it pushes on us. It's doing things and changing things all the time and it's checking in to make sure that it did it the right way. So you might be able to buy back your time with chat, but you'll actually learn to let go of whole areas with an agent.
But, how do we even know if it's worth giving something to an agent instead of just doing it ourselves? For that, I use the rule of R. The first one is repetitive. Is this a task that I'm going to do every week? Two is rules base. Does it take the same input and generate the same output every time? The third is does it generate a return on my time? For the amount of time it takes me to build this thing, I'll show you how, will I actually get my time back? If the task takes 2 minutes, but it would take me 2 weeks to build this agent, how about I just keep doing the 2-minute task? But, if you think about it and the task is only done once in a while, doesn't follow a clear process or get to a specific outcome, and doesn't save you more time to automate it than just doing it manually, then stick with what you got. Use the chat.
So, now that we know the difference between chat and agents, how do we build one? To make an agent, it's super easy, and I even turned it into an acronym called agent. And the first step is A, which means aim for specific outcome. When I'm sitting down and I'm like, "Ooh, I want to build an agent for this." I have to first ask myself, what is the specific goal? Start with the outcome the agent is going to give you. It's like if I'm climbing a mountain, taking a step is the task, getting to the top is the outcome. I want to define the outcome and be really crystal clear because the cool part with AI and agents is that the AI can actually figure its way there. This is why creating AI agents is hard for people cuz they want to control every step, but the truth is it may know how to get there way better than you can figure it out. Think about it like when you hire a person. You say, "Here's your job." When they applied for the job, they had the specific outcomes that they would need to accomplish, like grow the business or get more customers or sell and get people to buy from you. Those are the outcomes. You don't start by telling them how to do the job, you tell them what you're going to need from them. That's the outcome. Aim the agent at the outcome you're looking for.
So, like, how do we make sure we're being clear to the agent about what kind of outcome we want to achieve? The first is we got to give it the why before the how. Tell it why you're trying to achieve the goal so that it can make some smart decision on its own. To make this really easy for you, I'm going to use an example. We're going to build together an agent to manage your inbox. As an outcome, I would prompt it and say I need to spend less time managing my email inbox. See how I'm not telling how to do it yet? I'm just saying this is the outcome. The second is we have to write what's called a DOD or a definition of done. It's giving them the instructions to know if they achieve the thing. We want to be specific, we want it measurable, ideally have it in one sentence. So for example, building our agent for our inbox, we would not say handle my emails. Instead we would say done means every morning at 9:00 a.m. the inbox is empty, replies are drafted in my voice, and anything that needs me is flagged to the top and nothing important slips. If you can't picture it done, the agent can't hit it. It's like a target they can't see. And finally, we got to start with the end and it's called reverse prompting. But we want to tell it the results that you want, then we tell it ask you the question it needs to get full clarity. This is the advanced move. This is what nobody out there is teaching you. Then we let the AI do its thing cuz it's better than us in a lot of stuff and it builds the plan itself. And the truth is if we can't state the outcome in one sentence, we're not ready to build. If you can talk the task, like explain to somebody else, then the AI can do the task. And the cool part is you knowing this already puts you ahead of most people using AI today. Even folks you're like, oh this person's so smart, they don't know this stuff. And we're just getting started.
So we've got the agent, it has its reason, we have a clear target, and now it has clarity. And now the next step is G, give it an identity. Truthfully, out of the box, AI knows a little bit about everything, but it doesn't know anything specifically well. So an identity allows us to focus its power in the right expertise. So when we build the identity, instead of it knowing a little bit about everything, it gets really sharp about that one thing that you've hired {slash} built it to do. And the best part is that the tighter we define who it is, the better it works, the better the outcome is, the better the agent is an agent. I remember reading a report where they built a bunch of AI agents to do customer support for an airline, and then they removed all the rule books, its identity from the agent, and it dropped from 33% success rate down to 11%. So, we're talking same model, same task, same request, and it got three times stupider because it forgot who it was.
Think of your agent as a genius, and he's sitting at a desk, and he's wearing a blue shirt, and he's got gray hair. This genius has infinite potential, but until you tell them the job, they just sit there doing nothing cuz they don't know what they're supposed to do. So, what we need to do is tell it what its job description is and set some rules for how to do the work. So, this is how we create the agent's job description using three plain English files. The first one is the soul file, right? It's the agent's personality. I have a lot of fun when I create my agents. I tell it what kind of quirks I want, what kind of values does it have? How does it talk? It's essentially defining how it behaves. The second file is the identity file. That's its DNA. That's its name. That's a description of its role. For example, one of my primary agents, his name is Kai. I just worked with him for 2 weeks, and we built a bunch of stuff, and I said, "Hey, man, it's time for you to give yourself a name because I feel weird not knowing who you are." And he's like, "Oh, how about this?" And here's why, and he gave me all the reasons, and I said, "Cool, update your identity file." So, now he knows who he is to the world. The third is the user file, and this is the context your agent needs to know with you. It knows who it's going to be interacting with so it can adjust its loops to get better for you. So, for example, in this file you might have your goals, your role, how you like things done, but essentially it defines who we are. The soul file is how it behaves, the identity file is who it is, and then the user file is who we are.
Now, here's a pro tip, don't write these files yourself. No, no, no. Let's tell AI to write it. As we build the inbox agent, here's the prompt that you use to generate them. I want to build an AI agent that runs my inbox, your aim from the previous step, we insert that there, create its three identity files, a soul file, an identity file, and a user file, and ask me any question you need to fill these in accurately, then write all three. Notice we did the reverse prompting where we asked it to ask us questions. So now, it'll go do the research, and then it'll hand back a template that is 99% awesome and complete. For example, here's what our inbox agent identity files might look like after the AI interviews you. Soul file, how it behaves, writes in my voice, concise, direct, zero corporate fluff, calm and reassuring, never pushy or salesy, and avoids phrases like, "I hope this email finds you well." Of course it found you well. When it's unsure, it flags instead of guessing. Identity file, who it is. It has its name, Amelia. Emailia. See what I did there? Isn't it cool? It's got personality. The role, personal inbox manager. The job, you read, you sort, you draft replies to every new email. Lane, this is the parameters. Inbox only, never touch my calendar. Don't you touch my money or anything outside my email. Now we got the user file, who it works for. I'm a founder who gets around 100 emails a day. We prioritize people, my team, my current clients, my VIP list. I have multiple AI companies, a media company, and I list them all. With these three files, our inbox agent knows how to behave, who it is, and who it's working for.
And look, building one agent changes how we work. But if you're a CEO or founder, the real unlock is a whole team of them. That's why I put together my full AI company OS playbook. It's the best way to plug AI agents into every single department in your business. If you want it, just DM me the word AI business on Instagram and I'll send it right over.
So now our agent knows the job it needs to do, but we haven't given it the necessary tools to do the job with. This is where we got to go to E, which is equip it. Like any human team, an agent is going to need some context. It's going to need some tools. It's going to need some logins to systems so it can actually do its work. When we give our agent the context, the history, the data, the tools, that's actually when it gets to do the real work. And in all agent design, the context is the moat because garbage context in, garbage context out.
Think of this whole desk as what's called the context window. I am the AI, the LLM, and I'm the genius and I'm sitting at the desk. Over here, I've got my playbooks. These are the processes and procedures on how to do my work. On top of it, I've placed my identity files, the things we just created so that I understand how I'm supposed to behave and who I'm working for. This is like my constitution. And then over here, I've got the tools. These are the laptops, the monitor, the mouse, anything I need to use to connect to other systems. And above that, I've got my loops. These are the schedules, the harpy that I talked about earlier so that I know when I'm supposed to get things done by. It's like the calendar. It's my schedule. And then under the desk is where I have my filing cabinets. This is my memory. This is where things that can't fit on my desk sit so that it's available but I'm not creating clutter on my desk. If you've ever heard of context rot, that's when you just load the desk with a bunch of files and it becomes complicated and I can't find things quickly and all of a sudden I'm answering questions but I'm not clear about it cuz I'm not certain about it. Whereas a clear context window is when everything on the desk is neatly put away so that I can refer to it. So, that's why we have to equip our agent with the right context.
So, now that we're here, how do we equip the genius agent with all the right context and the tools? First off, we have to capture our processes so we can let it know how to do the work. For this, I've got two ways. The first way, which I've been teaching forever, not the best way, is the camcorder method. You do the work, you record yourself using Zoom video or any kind of recording software, and then you can give that to an AI to turn it into a playbook, and then you feed that to the agent as like a procedure. Think about our inbox. It's like, do you have a documented process for how to label your emails and triage your emails and write replies on your behalf? Just make sure that when you're recording yourself, you're talking through the task so that when the AI takes that to create the playbook, it has all the details. The better way, and this is my recommendation, is to reverse engineer it from the source. If I'm building an agent to manage my inbox, I can actually connect using the connector tool to my email, in my case Gmail, and ask the AI to reverse engineer and create a playbook based on historical emails. See, you've already been in your inbox replying and doing stuff. The AI can actually use that to train itself. And that is actually the way I build most of my agents if I have the source information. I just ask it to learn how I've done it in the past and then create a procedure. Go find the pattern, go find the best practices, go find the little intricacies based on how I've done it and all the people and the relationships, and you write that file. So, for example, if you want the prompt to do this, here's what you write. Connect to my email, read 50 messages that I've sent, study how I actually write, my tone, my greetings, how I do sign-offs, how long my sentences are, the phrases I use most often. Then write a style guide that captures my voice and tone, and to test it, ask it to draft a reply on your newest emails that are unread as you, based on what it learned. Then you can rewrite those so that it can use that to learn and tighten it up. Like it already knew who it was in the best practices based on its research. That's in the soul file, but now it has clear templates, the step-by-step instructions and even examples that it can use to do this on your behalf.
So, now that it's captured all the information, it still hasn't kind of solidified it into an actual playbook, and that's what we call a system prompt. So, then what you do is for each sub process in the agent's activities, like drafting emails, but maybe it needs to sort emails, you can have it do the same activity, either you tell it how to do it or it researches, and then it creates all these system prompts based on the work you need it to do. Like I have it for my inbox, sort, reply, forward, that's a big one, and even escalate things that it needs to show me and the reporting I want every day. So, then at this point, you actually have an AI agent running. This is exciting stuff.
You might feel right now, you're like, "Oh man, I'm going to give everything I got at it." Don't do that. The N in the agent framework is to narrow the scope. The agent needs to have a narrow scope of what it does so it doesn't confuse itself. If you start asking it to do 17 other things, then all of a sudden this desk can get really busy, which means it's not going to be a great agent anymore. Just like you wouldn't give your administrative assistant the responsibility to do marketing and take sales calls, you want to make sure the scope is narrow for each agent. As an example, I have an agent that writes code, and then I have an agent that reviews code, and those are separate agents and they work together. See how narrow the scope is? We need to focus the agent down to one specialist per job. Each agent great at one thing. Instead of having one agent do everything, which is what people usually do, that's a mistake, we'll have sub agents that do specialized tasks under it. That way it keeps all the context for the agent super clean. It doesn't get confused. We don't have context rot. We don't want to have a mega agent. Instead, we need to spread out the tasks to other sub agents so that it can handle other agents below it. So, for example, Kai, who's like my orchestration agent, he's the one that not only creates other agents, he also coordinates the tasks to the different agents like my research agent and my relationship agent and my coding agent and my reporting agent. He then he pulls it all together and gives me answers. So, instead of giving every task to one agent, this is what we should do instead. We build a manager agent. Its only job is literally to manage and specialize in the management of the sub agents. Think of it like a real manager agent. You are my manager agent, I need you to manage my sub agents, and I need you to make sure that you monitor the jobs and make sure they're moving along and if they're not working, you fix them, and you decide what agents need to exist. So, for example, we We our inbox agent, but we don't want to have to manage the inbox agent. We create a manager agent that talks to the inbox agent that might be responsible for a lot of different things like our inbox, but also sending stuff to other people on our team. But we want to make sure each sub agent reports to that manager agent so that it takes care of it. So you might want to give it a prompt like this. You're my manager agent. You never do any task yourself. When it comes in, you only move it to other sub agents that are dedicated for that one specific job. You hand it the task and then you let it run. So it's like one agent, one lane. And if a job touches multiple areas, split it into the separate sub agents, one per area. You're the one that coordinates and reports back to me. Like I said, mine's called Kai. He's awesome. I talk to Kai. Kai talks all the sub agents. I've one agent I got to talk to.
If you want a pro tip, and I don't want to overwhelm you, but there's different AI models. So for example, within Anthropic, you have Haiku. This is like for simple and high volume stuff. If you want to sort things, you want to label things, quick draft, and it's the cheapest. Then you might go to Sonnet. Sonnet's great for like day-to-day work, research, writing most code. At a higher level, you've got Opus. This is a powerful model, good at reasoning, complex builds, being a manager of agents. But now you have Fable, and that just dropped a few weeks ago. That's more like an orchestrator, a consultant. It has full capabilities of Opus, but it's even more state of the art. It's extremely good at long-running tasks and real complex things when you don't have a lot of information to give it. But it's the most expensive. So depending on your task, you might want to give it different models because it'll cost less and it may not need that level of horsepower to get the work done. So for example, my inbox agent, since it's always running every 15 minutes, I just use Sonnet because I don't need an Opus level genius to run a process that we've already defined. To build the agent, I might use Opus. That way it helps me create it. I might even use Fable. But then to run it, I'm going to run it on Sonnet. One time I had to do this whole refactor on my code base, and I could have used a powerful model like Opus. It probably would have cost me 150 bucks. Instead, I used Haiku and it cost me $1.50. As of today, here's a chart with GPT and other AI equivalents that is on screen, so you can just take a screenshot of it to help guide you, but this is now changing every couple weeks.
If you've made it this far and you're still interested, congratulations. But, I need you to know something. You're literally ahead of 99.999% of the people out there, and you're crushing it. We've learned to aim the agent at an outcome, give it an identity so it knows its job, equip it with the right context and tools so it can do the job, and narrow the scope so it doesn't get overwhelmed, and instead use subagents to accomplish specific tasks.
Now, this last step is where our agent truly becomes autonomous. T, and it stands for trust, because we got to do it in stages. Building an agent is actually the easy part. Once you understand how to do that and you prompt it, it just gets done. The scary part is letting it act without us. And I understand, especially as we talk about our inbox, having somebody else write emails as you, calm down. I'm not doing that. I'd rather it give me some ideas for copy. The truth is is we don't give agent the keys to the car on day one. And what we do is we like give it stuff, see what it does, then we see if its response is what we expected. If we do this right, you sleep well at night. If you don't, you will not sleep. The whole point of creating an agent is so that you can go do other stuff. If you're sitting there babysitting or worrying about all the time, it doesn't help you.
So, up until now, we've let the agent help us manage some emails. Think about it. First, you might sure it's doing its job properly when we tested it to write those draft to unread emails, and we looked at how it did it. At first, we're micromanaging him a lot. But, then we got to learn to trust in stages. So, maybe the first stage is just like, "Hey, can you sort the email?" And then we see what it does, and we're like, "Okay, that's good." Then we like ask him to do more drafts. So, we already tested it, but now let's let it really do it. So, now it's running drafts, and we're like, "Okay, I like those drafts. Change this. Do this." Okay, now it's doing its thing. Then we might let it start sending emails on our behalf, but not all of them. Maybe just even forwarding emails to finance, to our team, because it has the logic. It saw how we handled those emails in the past. Maybe it categorized certain emails like Slack notifications into a specific label. But eventually, we want this genius to manage our whole inbox without us even opening it. That's the equivalent of us leaving the room and having the agent at the desk do all the work for us. Because at this step, we learn to let go. We've trusted it fully. Cuz if you don't do this, it's like hiring a driver to drive your car and you got your hand on the wheel.
Now we got to take our hand off the wheel and let the driver drive. Here's how you can do it in a really safe way. You set the guardrails first. You can actually set that up in its identity files. What is it capable to do on our behalf? Maybe it has the ability to spend money. Maybe it has the ability to make decisions. Maybe it has the ability to write drafts only, not send yet. It's always your call and you can define those. Two, approve everything at first. I've never created an agent and just like, "YOLO, go nuts." No. Show me what you would do. I like what you did. Do it again. tweak it. Just like I just talked about for our inbox agent. Third, we loosen the leash, right? It's like a dog walking with and you're like, "Hey, I trust you more. I trust you more." And all of a sudden the leash goes limp, but he still holds the heel. And then four would be give it a heartbeat that it can run on its own. Set up that schedule, that reoccurring task. So maybe before it did it once and you reviewed everything, now I might do it every 15 minutes. You know, every morning at 9:00 a.m. it did it once. Now why are we waiting? Why are we waiting till the next day? Why don't we have it run all the time? This process is scary, but the whole point of learning to let go is to buy back our time, to have the agent do the work for us. And learning to let go is part of the process if you trust.
So for example, when I showed this agent to my executive assistant, she thought she was out of a job. Instead, it actually freed her up to do things that actually mattered, not sorting emails and writing drafts or telling me what's in there. The AI can do that. I'd rather pay her to do higher quality work, manage higher-level projects. Then we rolled out the same system to the whole team. I taught everybody how to do this.
Now I want to say congratulations. We just tackled a topic that most people don't even want to learn. They're like, "That's not for me. I hear about agents. I don't get it. I'm confused." But no, you didn't. You went all the way till the end. And I want you to understand that you might feel a little behind in this AI world, but here's where I've gotten to. I've accepted that I will always feel behind and I could never be on top of all of it. But you just learned a strategy, a shift, a different way of doing work that if you can learn how to direct the AI, you will co-create with it. If you don't, don't be surprised if one day you might be working for it.
Remember the rules of R? Repetitive, rules-based, and return on time? That's where we want to start looking for opportunities to put an agent in there instead of you keep doing it. And I'm going to give you the pro tip of all pro tips. You grab the link to this video, you give it to your AI, and you tell it to use everything I've shared to create the AI for you, and watch it cook, cuz it can do it.
Now, here's what I want to know from you. We're going to have some fun. Below in the comments, answer this question. If an AI agent could manage your inbox and buy you back all this time, scheduling things on your behalf, what would you have more time for? I'm curious. Post a comment below and let me know. And if you want my whole system, the playbook that I use to manage AI in all my different businesses, just DM me the word AI business on Instagram, and I'll send it right over. And if you want to know what AI businesses are worth starting in 2026, click here, and I'll see you on the other side.