Transcription
I have this organization of 10 different OpenClaw agents doing work for me 24/7, 365. It literally creates me value all day and even when I sleep. I'm literally the CEO of my own AI company. In this video, I'll show you how to do the exact same thing. I'll show you how to build your own AI organization, determine which agents do what, and which models you should be using to save tons of money. By the end of this video, you'll be your own AI agent CEO with an army of AI agents creating value for you at all times. Let's lock in and get into it.
So, what you're seeing here, I believe, is the future of work. Every person having an army of AI agents working for them. I've had them going for weeks now, writing me newsletters, building apps, scripting videos, finding trending topics, getting me news the moment things break. I've been working a hundred times more efficiently because of this company working for me. These are all open claw agents and sub agents just continuously around the clock doing things.
What we are going to go through is how you can set up this same exact team as well. How to determine which agents sub agents to use, what they should be doing, and help you figure out some use cases that will be helpful for you. If you're watching this right now, you probably only have one open claw going. And that's totally fine. 99.9% of people have the same exact setup, just one open claw that you kind of talk to and it does things for you. But the ideal state of Open Claw is what you see here. Having multiple agents reporting to each other and having an entire hierarchy of sub agents that report in and do work for you.
So the first thing we're going to do as we build out this AI organization is determine what we want our organization to look like. What sub agents we need, what agents we need, what models they need to run on, and how you're going to communicate with all them. This will help set it up so you're getting tons of value out of your organization. This is going to help you get way more work done, automate way more of your work, and just make a lot more money.
So, you have your open claw right now. I call this my chief of staff. This is Henry. This is running on Opus 46. But what we want to do now is determine who our first employee will be under your open claw. A lot of people try to rush in and they try to just build five sub agents right off the bat that do all these different things. That is a massive mistake. You want to start one step at a time. We want to find one sub agent that does one use case very well and then we can layer on sub agents after. It's like the people that run off and buy like 20 Mac minis to start. It's a mistake. Start with one. Start very small and narrow and then we layer on after.
But how do you know what your first sub agent should be and what their role should be and what model they should be using? That depends entirely on what your day-to-day tasks are. If you're a coder, you should probably build a coding sub agent that does work for you. If you're a content creator, you should have a writing sub agent. If you're a researcher, you should probably have a researching sub agent that does work for you 24/7. If you're not sure which one to go with, I recommend setting up a coding sub agent. This is a sub agent that's spun up anytime you need an app built. And it doesn't matter if you're a coder or not. You should be building apps with your Open Claw at all times. It is really easy to do. You don't need coding knowledge. So if if you don't know which sub agent you want to build right away, let's start with a coding sub agent.
Real quick before we create that sub agent, we're going to want to determine which model it runs on. This is a big advantage to setting up a organization like this is you can have the right model doing the right task at the right time which will save you money. If you do everything with Claude or everything with Chat GPT, it's going to be very expensive. So for orchestration, the best model is Opus 46. For coding, which is going to be the agent we set up next, Chat GPT 5.4 is going to be the best. It's going to be the most impactful and also be incredibly cost-efficient. Doing everything with Opus is just going to be wildly expensive. And then research and writing, you can go with a super cheap model here. You can go with one of the Chinese models like Kimi or Quen. I prefer local models. We'll get more into that later in the video on local models and how that fits into this. Or you can even do Open Router and choose like a really cheap model on Open Router. But we're going to go with Chat GPT 5.4 for this because we are going to set up a coding sub agent. Then I'll show you how to set up other sub agents after like researchers, writers, things like that. But I still recommend Opus 46 as your orchestrator. If you got kicked off OAUTH, then I recommend Sonnet as your orchestrator and just using the API, which shouldn't be too expensive. But this is huge to know and huge to use because if you use the right models for the right tasks, you're going to save so much money and get way better results. Don't just use one model for everything.
So, I'm here in my orchestrator agent. My orchestrator agent was spamming me all day and night over a coding task that kind of crashed, but we cleaned that up. Let's talk about how to spin up this sub agent. This sub agent is going to be our first employee. So, we're going to say this to our main orchestrator, Open Claw. So, here's the prompt. I'll put it down below in the description. Feel free to pause here and paste it in and work alongside with me. It's probably the best way to consume these videos. I want to build a sub agent that you will use for all coding activities. Anytime you need to code something, you will use this sub agent. Please name the sub agent JSON or whatever you want to use and have it run on Chat GPT 5.4. Let me know what we need to do to set up the OAUTH or API key.
So, this is going to do a few very interesting things. This is going to put in your Open Claw's memory that every time it needs to code, it's going to spin up this sub agent. It's also going to assign a model to this sub agent. So in the config of Open Claw, you don't need to worry about it. It assigns models to all the different sub agents it uses. So it is going to assign Chat GPT 5.4 for that, which I think at the moment is the best coding model and also very cost-efficient. And then it's also saying, hey, help me set this up. So it either asks you for an API key or gives you the link to OAUTH or walks you through how to do that. Hitting enter on this will have your Open Claw set it up. You never want to set up things on your own. You always have your orchestrator agent do it. If you go in and you're changing config files, there's a very good chance you'll break your entire Open Claw. You don't want to do that. Just have Open Claw configure everything itself. You hit enter on this and it is going to set up your sub agent that we used for every single coding task.
Once you do that, you can test it out. So you can say, "Hey, build me a quick snake game app." Or, "Hey, build me a quick 3D first-person shooter." And it will say, "Okay, spinning up JSON to code that out for you." And you can confirm that it worked. You now did two things. You made it so your whole entire setup is more efficient because you're using sub agents. And I'll go into the advantages of sub agents in a second. But you're also saving money because you're not using Claude for every single task. Claude is by far the most expensive model. So if you're using Claude as your main orchestrator, you don't want to be using it for everything. You're saving money because you're using the right models for the right tasks.
I usually get this question a lot. Should I be setting up multiple Open Claws? Should I be setting up separate sub agents? What's the best model? There's so many advantages to sub agents. My setup is kind of complex. I have Open Claws. I have sub agents. I'm keeping this beginner-friendly for now, but right now sub agents are a big advantage because one, parallel work your Open Claw can spin up to five sub agents at a time. So that's five sub agents that could be doing work at the same time. So it's getting work done faster. You're saving money because you can use again the right models for the right tasks. Each one of these sub agents gets its own context window. So if you have your main Open Claw doing everything, it can get confused and slow down because it's having context in its context window about every task it's doing. Sub agents keep the context windows clean. Failure protection. So if any of the sub agents ever crash, your orchestrator agent can just figure out what's going on and fix it. If your main orchestrator agent ever crashed because you have it doing 20 things at once, everything's toast. So, it's good insurance to be using sub agents. And then auto archive. So, when agents are done, when sub agents are done, they can just get shut down and cleaned up so things keep moving smoother. So, that's why we're setting up sub agents here and not multiple Open Claws.
So, we got our first sub agent set up. They're doing some coding work for us. Next, let's add some things to our mission control that make this all a lot more organized. If you haven't set up a mission control yet, I'll put a video down below for setting up a mission control. I just filmed it last week in Japan about how to get this set up. You should definitely see that if you haven't yet, you can watch that right after this. But I want to show you a few things you can set up to be better organized here.
Number one is this team screen. This is an org structure. I like setting this up once you're spinning up sub agents because this does a couple of things. One, it keeps you organized, so you know which sub agents are in charge of what, but two, it also keeps your orchestrator agent organized. Henry, at any point when it needs to do work, can come in here and check to see which sub agent it needs to give work to. Henry's the boss. I don't tell it which sub agents to use. It figures it out on its own. And because I have this team screen, it can figure out on its own which sub agents it needs to give work to. So, what you're going to want to do is you're going to want to go to your main orchestrator agent and you're going to want to say to it, I want a screen in our mission control that shows our entire team structure, showing each sub agent, each agent, what model they're working on, and what their responsibilities are. This will keep you super organized, and you'll be able to check at any time like this as you scale up. It's a lot easier to scale up and add more sub agents when you have a clear picture of what these agents are.
In a second, I'll go through setting up a third sub agent, fourth sub agent, and show you how to give them tasks that are really, really helpful. But I want to show you a couple more screens here. And real quick before we do that, every Friday, by the way, I run a live boot camp in the Vibe Coding Academy. I go through my mission control. I go through my sub agents. So, you have any questions for me, feel free to join the Vibe Coding Academy down below and you can join the live sessions and ask me questions live. They're a lot of fun. A lot of people enjoy. There's over like a thousand people in there now. But anyway, back to this.
This is the calendar view. This is an important screen to set up as well. This will allow you to track every scheduled task your sub agents are doing. So, one of the great things you can do with your sub agents, again, I'll show you how to do it in a second, is schedule tasks with them, right? So, I have different things on here. I have a sub agent that looks at trends every day at noon, see what's trending on X, on YouTube. I have sub agents that write me morning briefs. I have sub agents that write me scripts for YouTube videos. And I can come in here in this calendar screen and see every task that is scheduled, every cron job that's scheduled by every sub agent. So you want to set up a calendar screen. Again, very, very simple. You go to your Open Claw, say, "Hey, I want a calendar screen in my mission control. Please build that out that shows all the scheduled tasks and cron jobs all our agents and sub agents have." You hit enter on that. Boom. You're good to go. You get something that looks like this. If you want, by the way, you can pause at any time, screenshot this, and actually give it to your Open Claw and say, "Hey, buddy, build this." And it will spin up your new coding sub agent to build this out. So, if you think these look good, I spent a lot of time making sure these look good. Feel free to screenshot it, send it to your Open Claw agent.
And lastly, this is my favorite screen. There's not much utility to this screen, but it is a lot of fun. And hey, you're allowed to have fun when using AI, and this is the office view. This allows me to see each agent working, each agent walking around doing different tasks for me in real time. It's a good way. I just have this up as I'm working during the day and I can peek over, oh what are my agents doing? What's Ralph doing? What's Codex doing? And I can see that all happening here. So, you should also be building out an office screen like this. It's just fun and you're allowed to have fun. Feel free to screenshot this. Give it to your Open Claw. Say, "Hey, can you build me a 2D pixel art view of all my agents working? I'd love to track my agents doing work." Give this to them and you'll have this really cool screen set up.
So, you got your coding agent set up. Let's set up a couple more that would be helpful for you here and then I'll show you how to get really custom and figure out other agents you need. But, let's talk about a researcher. Let's set up a researcher sub agent. Odds are you care about something going on in the world that's current. If you're watching this video, if you watch my channel, you probably care about AI and current AI news and trends. So, you should set up a researcher agent. Everyone should have one. This researcher agent for me goes out every morning and finds what's trending in Open Claw and Vibe Coding.
So, here's what I like to do. I want to build a sub agent that you will use for all research activities. Every morning at 8 a.m., please have the sub agent search the web to find trending news about AI and Vibe Coding and AI agents. Also, please put them in our AI agent org chart in mission control. Please use Quen 3.5 for this sub agent or whatever model you want to use. I think the cheap Chinese models are great for this activity, but it's up to you what you want to use. I personally use a local model that's local on my computer running 24/7, so it cost me $0, but that's up to you. You hit enter on this, and now you'll have a task scheduled every morning at 8 a.m. where your sub agent will go out and find trending news for you. I like this because it helps me stay on top of what the newest AI tools are so I can use them, so I can show it to you guys. If you want more helpful AI videos like this, feel free to hit subscribe and turn on notifications down below. All I do is make amazing AI videos.
So again, this is now going to set up a scheduled event for your agent. Every morning at 8:00 a.m., your agent is going to go in and they will be doing these tasks for you. So, as you can see here, if my trend radar is at noon, yours could be at 8 a.m. whenever you want. And now you can look at your calendar screen and see exactly what your agents are up to and when they have tasks coming up.
Now you can start to see how this organization is coming together, right? We set up some basic agents, right? If you set up two, you're good. You don't need to have 20 like I do just now. You have them going, you see the results, and as they go, you think of more use cases you can have them do. And you slowly layer on top that. You don't want to do a thousand at once. I hate these videos where the creator goes, "Oh, you got to set up this army of 50 different sub agents. Here's a GitHub repo of 50 different sub agents that will do work for you and like no that is I think that is the biggest mistake because it's going to overwhelm you. You're not going to know what the hell is going on and it's going to turn you off and cause you to use none of it. The best way to approach this is one agent at a time, one sub agent, one coding sub agent. It builds you some things. Then you add on one research agent. Then from there you can start doing really interesting things. You can have a thumbnail designer. It doesn't have to be a thumbnail designer. It could be just any sort of visual designer you want. This could be powered by Nano Banana. And every time you do something, maybe you find a new story, maybe your content creator, your researcher agent finds a new story, it hands it off to your visual designer and that creates a thumbnail for something or a visual for something, right? We start to slowly add to the organization. Maybe you run a business and you have an X account, a YouTube account. You want a sub agent to manage those things. You slowly layer on another sub agent, a social media manager. You do this one step at a time. I wouldn't add more than two in your first week. Master those two, get them in your workflows, and then you can layer on from there.
Let's talk about customization real quick. I showed you some basic use cases, coding, research. What it comes down to is you want this organization to be custom to you. And the best way to do this, again, if you watch my videos, this is going to be like the 10,000th time you heard this. Reverse prompting. Reverse prompting is when instead of you telling what the AI to do, you ask it what you can do based on what you know about me, my goals, my workflows, what would be some good sub agents we can spin up and assign. Your Open Claw knows way more about you than I do. So, it will know based on all your conversations which sub agents you should set up. It'll give you a list. Choose just one, have it set up, and start using it. This is going to be the best way to set up your own AI organization. Feel free to copy me. I don't recommend it because me and you are probably different people with different workflows, but you can copy me to start, but always make sure you reverse prompt so you can set up your own custom organization.
Let's also talk about future state real quick here. What I think the future of all of this is, and I think the future of all of this is local AI. Right now, if you have sub agents going 24/7, 365, it could get pricey. It could use up a lot of your limits. That's why I think local AI is so important. This is my setup here. I have three Mac Studios, a DJX Spark, and a Mac Mini. You don't need to do this. You don't need to run out. This is like $40,000 computers. You don't need to do that. You can start small with one computer, one DJX Spark, right? Load a model onto it and have that model handling one of your sub agents activities. Maybe you load up Quen on there and then Quen 3.5 handles all your research. Now you've offloaded one specific task to a local model that will cost you $0 in tokens. Yes, you're spending some money upfront for the computer, but now in perpetuity till the end of time, it can do work for you 24/7 without costing you anything. This is what I believe is the future. Having swarms of local agents constantly all day and night doing work for you because it's going to cost you a little bit of money upfront, but no money after that. If you were to swarm a hundred sub agents doing work 24/7, you're probably paying thousands and thousands and thousands of dollars a year. If you buy computers, it's just one upfront cost and then it costs you nothing after that. This is what I think the future state is. I'd recommend thinking about this and think about how you can slowly layer this into your workflow. But this is what I believe the future state is. Hey, I'm open to disagreement if you want to let me know that I'm stupid. I don't care. But this is what I think is the future and this is what I'm investing in.
I really hope this was helpful. In the next video, I'm going to talk about building your own AI research lab, how you can have your sub agents actually training models for you. It's a little advanced. If you think that's too much, let me know down below. If you want to see that, let me know down below. It's a little bit outside of what I normally do, so I want to get feedback before I film it. So, let me know down below. If you have any other questions for me, I'm always in the Vibe Coding Academy. You can post questions in there. There's also a thousand people to answer your questions, too. It's a whole community of Open Claw and Claude Code people. Link for that down below. Make sure to join. Hope this was helpful.