Complete Agent Zero Tutorial (Agent Zero Guide For Beginners)

I have spent the last year of my life searching for the ultimate open source AI agent framework. From Eliza OS to Crew AI to Pedantic to Archon, you name it, I've tried it. The first one that actually works out of the box like advertised is Agent Zero, and I finally feel like I'm coming home. This is where I'm going to build and learn and innovate because it's super easy to use and super powerful for a non-developer and non-coder like myself.

Agent Zero is fully open source, but there's no vendor lock-in, no subscription fees, just raw capability that you own on your computer. So if laws change tomorrow, if Anthropic raises its price by 10x, whatever, you can have the power of open source intelligence at your fingertips. It runs locally in a Docker container. It can't mess up your computer by deleting important files like Cloud Code. And it gives you full granular control over your agents, which we're gonna get into in this video. It is super incredible. How much control? So much control that if you don't put in time to making your agent do exactly what you want, it will do other things. It is on point with the directions you give it. It is a great reflection of your own consciousness surrounding what you want to build. If you do not know what you want, you will not be able to build it. Agent Zero will help you build what you know you want. This is intention-driven development at its finest. We're going to touch on that in this video as well.

In this workshop, we'll run through getting Agent Zero set up on your machine. This is a total noob guide. It's easy. We'll connect it to your preferred AI models, and we'll build a custom sub-agent that calls its own sub-agents that spawns its own sub-agents. And by the end, you'll have an understanding of how to build your own completely sovereign AI system that runs locally on your computer. So let's build, baby. So remember, there's time codes down below. If there's a section that doesn't really interest you, feel free to skip ahead.

We're going to start off with some words by Jan Tomasek. He's a founder of Agent Zero. He joined us for this community call at the beginning of the workshop and introduced Agent Zero from his perspective. So I'm going to include a little of that here, and also just the overarching perspective of why open source AI agent frameworks are so damn important in our world today. And from there, we'll hop right into installing Agent Zero and using it.

We focus on transparency and open source. And that's what I think is the most important part right now, that you have something that you can download from the internet, and you can run it locally without anyone knowing it, without any paywalls, without any logging on remote servers, without anyone sniffing on your information, etc. So I would like Agent Zero to become a platform for agent builders. Something like Linux is for developers. I can't wait for the time when there's like Agent Zero templates on GitHub. So here's an Agent Zero, just whip it up, deploy it. And it's like your virtual assistant, right? Except it's an agent or it's your business strategist. And then all of these templates, I imagine will be really refined with their own sub-agents and have whole teams. And all of it is open source. And so I hear you completely why it's so important for this to be out here right now as an open source tool because yes, other companies are building these technologies, but very little of it is open source in such a way that, like you said, you can spin this up at home locally. No one knows you're doing it because who knows what will happen with the politics of this technology in the future and say, oh, hey, you know, if you want to use OpenAI, you have to use WorldCoin and give them, give Sam Altman everything about you. And there needs to be an opportunity to opt out for the average person to not have to play that game. And I even see how dependent I've become on Cloud Code for a lot of my AI usage. And I prefer Anthropic to OpenAI, but I still would ideally not like to deal with either of them.

My experience with Agent Zero completely changed when I started using proprietary models instead of open source models. The performance is way better when I use Claude Sonnet instead of Llama 3 with Ollama. It's just way better. And so we can't hide from that. So we just have to be creative and clever in ways that we mix and match. And we make sure when we do need privacy, we have open source local models available to us. But most importantly, and this is what the AI Captain School is really all about, it's about claiming sovereignty over the foundational technology. So as all this stuff evolves and changes, you can tweak it and you know how to tweak it on your own. So you're not beholden to these big tech companies. So like Cloud Code, it's pretty cool. It's agentic. It can spawn new agents and stuff. But it still feels like a black box that I can't see out of. And I can't just say all the time. I can say modify your behavior and it'll edit CloudMD. But a lot of times it doesn't even reference CloudMD because of whatever reason. Whereas Agent Zero, it's like built into the code for it to reference its memory. So it can't escape it. And so just little things like that, where it's like, okay, well, that alternative doesn't serve me. So I want to use this instead for these purposes, because it does serve me. It doesn't forget to check its behavior file. And I say that a lot of my videos, because it's been such a big deal for me. I'm like, why am I putting all this time into my CloudMD file, if it's not going to fucking reference it? But I know exactly what you're talking about. And I have the same experience with Windsurf, unfortunately. Like it completely ignores my prompt files. Yeah. And Cursor did a pretty good job. I think Rue code does a good job. And it's still all just like prompt hierarchies, right? That's really all it is. So I don't even understand why these aren't working, right? Like it's all agents are just hierarchies of prompts, essentially. So why isn't this working? Anyway, when it comes to Cloud Code, it's great at coding. It's great at building things. It's very agentic like that. But do I trust it or imagine it will go in the same direction as Agent Zero? Or it can self-code its own new agents it needs to accomplish tasks? Like, yeah, probably. But I probably won't have a say in the granular detail of those sub-agents that it creates when it wants to. But with Agent Zero, I can have full control over it. And I think that's the heart of the matter. I think it's really still early for the average person to understand how important this is. But in just a few years, five, 10 years, it's going to be really clear why the average human being that's not an elite should be able to have access to self-spawning AI agent software, essentially. I mean, we'll come up with a better term one day, but that's my perspective on the project. And I love it. And thank you for all your hard work, Jan. Thank you very much. And thanks for coming to say hi and introduce it. And yeah, I know you got to go. So I'll talk to you tomorrow. Okay, thank you guys. See you guys. See you.

Okay, thank you to Jan for introducing us to the project. This is what my mission in North Star is really all about when it comes to creating AI agents and learning this technology is we really have to be aware that these big tech companies do not want what's best for us. They want control, influence, whatever it might be. So to that end, that's why the AI Captains is here. You can also be learning more about Agent Zero and the Agent Zero School. It's completely free. It's a big community here. I post in here sharing what I'm learning and questions I have. Because if you've never been part of an open source project and community before, they are, as Jan kind of said, all the code has to be really clean, but also like anyone can contribute and become a part of the community of contributors and send in GitHub pull requests and just be part of it. So if you have a really good idea, you could send a pull request like these 89 here. And if the team, the core team says, hey, this is a good contribution, we're going to add it to the code base. And so that's what happens. And then you see there's an issues little forum here in the GitHub. And that's really what GitHub is all about, I would say, from an open source community perspective, if you've ever wondered. It's not just a place to host your code repositories, but it is a place for just communities to build projects together.

So Agent Zero recently was number one repo of the day on GitHub. It was trending. And once again, it has a strong community. And it is just a really cool project. I was telling Jan before we started recording, I was looking at AI agents all year. Okay. So I started in January. I started with Eliza OS. I went into Crew AI. So many GitHub repos I have explored as AI agents. I've explored, tried to get them working. And Agent Zero was the first one where it really is an easy installation. And it does work basically out of the box. There are a couple. You don't even really need to use the command line, actually. There's two things you need to do to keep it persistent running on your computer, which I'll show you today. But the installation is super simple, and that's where we're going to start here. I'm going to walk through the installation. I'm going to walk through the interface here. I'm going to show you why I'm getting this error, and that's because of Ollama. So we're going to switch to Ollama there in a bit. We're going to walk through it, how it works. I'm going to show you how you can get Ollama local models on there. I'm going to show you how to find the models you want to use if you're using Ollama or OpenAI or whatever, how to make sure your laptop or computer can handle those models if you're doing it locally. And then I'll show you some kind of side projects that are being built off of Agent Zero. This is like a VTuber that's built off the Agent Zero framework.

So yeah, the installation is super simple. You go to Docker Desktop, docker.com, okay? And then you find Docker Desktop. Download Docker Desktop. Download it for your computer. You'll get an installer file. You install it. Easy. Then after it's installed, you've dragged and dropped it. You've launched it. Go to your advanced settings and make sure you allow the default Docker socket to be used. That is important here. Then you're going to run Agent Zero. So in order to do that, we need to pull the image. So in Agent Zero, excuse me, in Docker Desktop, there it is, we are going to go to images. And what we're going to do is find the agent zero image and then run it as a container. So if you've never used Docker before, Docker basically creates containers for software on your computer or server or whatever it's running. So if you have a server that has Node.js and this one, say you want to run multiple scripts and it needs multiple libraries like Python over here, JavaScript over here, C++ over here, whatever, then it's ideal to use Docker containers. So each thing, Python over here stays in a container. Node.js stays over here in a container. And that is how you can kind of maximize the usage of a server or your computer for that instance. But in the case of Agent Zero, the Docker container provides an environment where it can have full control and it won't fuck anything up on your main computer. When you're running Cloud Code on your main computer, you have to give it permission to run commands because it very well could be running a command to delete everything because it has access to everything. We don't want this, especially as the internet goes more like the AI internet goes more rogue in like kind of a corporatocracy dystopian direction. We don't really want to give them any kind of AI agent full control. And maybe right now it's fine, but maybe in a few years it won't be fine. So Agent Zero operates within a container. So when you have a chat going on and it has to, whatever, run code, which it had to do here so I can install my Cloud Code MCP server, then it can just do all this stuff as a pseudo root user with full permissions. It can just run all this stuff and we don't have to worry about it messing anything up on our main computer. So I hope that makes sense.

Docker, we have Docker here and we're going to search for agent zero AI forward slash agent zero. This is the name of the, we'll call it the repository. We click that and we click run. Now, I already have it installed here, so it might look different for you. Let's see what it looks like on here. Yeah, you want to pull it. That's what it is. You're going to pull it first. And once it's pulled, then you can run it. Now, I don't do this. It's going to become easier in newer versions. But I would just say skip this if you're a newbie, because that will become easier in newer versions of Agent Zero. Jan just told me. But now we want to run it. So we've pulled it. And once it's pulled, we're going to click run. Okay. Now it doesn't end there because it's going to open up this. Look, I can actually run a second Agent Zero container. We'll do that. I'll call it agent zero two. And we want to give it a port. Okay. So the port here can really be anything. But if we just put zero, it will assign a randomly generated port. I'm just going to put zero there. And we'll say run. Uh-oh. We've got a problem there. Bad parameter. Invalid container name. Okay. That's because it should probably be with no spaces. agent zero two. And then we'll just do zero there. Run. Boom. Now you'll see, Axe, you can't see this on your screen. Sorry, everyone in the school call. But I've got Docker Desktop now. We see the terminal and it's running. It's actually going to look pretty much just like this. So the container will start. Alternatively, you can also run this stuff in your command line. Okay. You can literally just run it that way. But once again, as new people to Docker containers and this kind of thing, it's a lot easier to just follow these directions with the graphical user interface.

Now that we have it running, we can go to our containers tab and now we'll see, I have agent zero two here. So I have agent zero one up there and that was just given a random name magical Clark. But now we have agent zero two. So I can click that now localhost port 55001. So I click that boom, we have a fresh install of Agent Zero. Now notice how it looks a lot different than this install of Agent Zero. Let me zoom in here. That is because one of them is the Agent Zero I've been using and one of them is the brand new Agent Zero we just installed, which we'll doesn't have anything in here. I don't have any API keys, nothing. So I'm going to show you what this looks like on a fresh install. And then when we go into actually using it, we'll see what it actually looks like when you've got everything set up. So the first thing we want to do is go into settings and we start here at the agent settings tab. And we'll see that we have this little menu of things we want to configure and it'll take us there. Okay. That's what these fun buttons are. Here we see our default agent profile. Now, default is just default. You can make Agent Zero the default and it has a certain way it runs. You could make your content creator agent the default, which is what I made. I have a school optimizer for content I want to make for school. And then in the, if we go to the new one, you'll see the default models here. You have a hacker, which can help you with cybersecurity. It can help you test out your vibe coded projects. You have a researcher, you have a developer who excels at coding, and then you have a subdirectory of knowledge. And if you want, you can create another folder here.

Now, let me show you real quick what Agent Zero allows you to do here. It allows you to look at its files. So if we click the files button here, then we're in the root directory. Let's go up. Let's click A0, agent zero. And here we see agents. We see docs. We see instruments. We see knowledge. This is the default knowledge base folder. So when you want to add documents for knowledge that you want your AI, your Agent Zero to reference, like for your business or for your website or for whatever it is, you want to put it in here. You can either upload it or you can simply attach it and say, add this to our knowledge base. And it'll throw that into the files. In fact, let's do that. Great. Here's a strategy document. Now, it's not going to work because I don't have anything set. I don't have an agent set. So let me show you how to set an agent. And then we'll go back there. So chat model provider can choose what you want here. I would say Ollama is a great one to use for anything that's not coding. And then in that case, you want to make sure you've got the right model selected, right? And if you're using Ollama, you can get your API key on their interface. And then, yeah, Llama 3 8B is what I'm going to use here. I'll put that in here. And I'll even go off screen real quick. And I'll get an API key so this can work. But, of course, you can also use any. You can use Anthropic. You can use OpenAI. You can use Grok, XAI, and any other OpenAI compatible projects, which just have a different base URL than OpenAI's. So let me generate a key here. 802. All right. And so now, once we have that provider selected, we're going to want to go into External Services and put in our API key. So we can just put in our API keys for every service here. I am going to move that over there to not dox my API key. I'll paste that there. Okay.

So now here in the settings, we'll see that here in External Services, it'll say show, okay, I've got an API key set. Let's go back to the Agent Settings. All right. Now I can put that here. We can put the context limit, which I believe. Oh, 262,000, which is quite a lot. So let's put 262,000 here. And then it allows us to tell us what context window space we want to keep for the chat history. And the rest will be optimized to fit it. So like the point three that's remaining here will be used for the system prompt that Agent Zero uses for the retrieval augmented generation of the knowledge and memories and the response it makes. So it will continue to compact the conversation and save it in the memory, which will look in a little bit. And you just tell it what the context length is so it knows what size everything needs to be to perform optimally for your tastes. You've got some other things here, like does the model support vision? I'm not sure if this one does, so I'll just Google it. Does it support vision? Not even Google, Brave Search all the way, baby. Does not support vision. So we'll shut vision off there.

Next, you have your utility model. Okay. This is a smaller and cheaper, faster model. That's really about not responding to your, not giving you responses or really thinking, but doing things like organizing memory, preparing prompts and summarizing. So what is stuff that like any AI can do. Well, they're really good at looking at patterns and spitting it out back at you. They're really good at taking really specific directions and doing long tasks, right? So it's not so agentic as much as it is, I don't even know what the adjective would be, but just like old school AI, right? Like from a few years ago where it could just get tasks done for you, but it wasn't super smart or creative or clever. So this model, you want to be more smart or creative clever. And then down here, the utility model, it's really just a smaller, cheaper, faster model. Now, let me look at Ollama's model list again. Do I want to pay that much? No, let's use this one. OpenAI, GPT, OSS, 120B, $0.07 per million token input and $0.30 output. That's super cheap, which means it's probably fairly fast too. So I'm going to put that as my utility model. Boom. Then you have some more settings, requests per minute limit, input tokens. So these limits aren't too important for now. One day you might be so advanced where you do want to add additional parameters like the temperature or the top P, et cetera. Not a huge issue for me right now.

Then the web browser model, you want to make sure that this agent can browse the web. So I would just use Anthropic and I would say Claude Sonnet 4.5, right? Or I could do Ollama and I think Ollama Medium up here. I'm pretty sure that has vision too. That can surf the web. And then you can make sure it has vision. So actually, yeah, we'll just use Ollama and we'll say Ollama Medium Mistral 7B. So this is why I like Ollama. Because you have all these different open source options to choose from. I like Ollama for that reason. And that's why I'm always using it. And it's pretty cheap. And if you stake their tokens, you can get recurring API access. So every day you get, you know, however many dollars worth of API credits, and it starts over the next day. So you're not paying for it all the time. You make one upfront investment, and then you get API credits forever, which I just think is awesome. But you want to make sure you're using those API credits efficiently, which means you want to use the right models. And Agent Zero helps with that because you have three different model options just for three different tasks that it might need to do in a conversation with you. Oh, and then there's the embedding model. No need to change. We can just use this Hugging Face model here. No problem.

Memory. Now, this is a cool part of Agent Zero, which I'm still learning. But basically, there is a memory folder and a memory dashboard. And essentially, there's auto recall, not total recall, auto recall. And actually, it is kind of probably total recall now that I think about it. But based on the context, it'll look into its memories, and it'll remember everything you've been talking about or what it's been learning with you or this conversation in particular, what it is. It keeps everything in there and compacts it. Now, if I show you the dashboard here, we haven't... This is our new one, right? So we actually don't have any memories here. But if I show you the memories in my main setup, in the memory dashboard, we can see the metadata of what the memory is about and then what actually happened in the memory. So in this fragments category, in a conversation, user instructed the assistant to never code directly for any coding tasks, Python, JavaScript, etc. The assistant must always use the Cloud Code MCP tool. And this rule also applies to any sub-agent. So this was my solution to open source models not being very great at coding. Cloud Code excels at coding, and it's so good. So I installed the Cloud Code MCP server in Agent Zero, and then I said, okay, Agent Zero, whenever you need to code something, you use Cloud Code MCP. This will also save money compared to using just the Anthropic Cloud models in the model section, because then you have to pay API call by API call. If I were just to set this to Anthropic, you're not using your subscription. Whereas with a Cloud Code subscription, you're just paying monthly. You know how much you're paying, and that's why this model works. So we have all this information about the memory. We can also delete it if we want. Here's a solution memory. And this was just how to get Cloud Code MCP working with the right wrapper inside of a Docker container, basically. So this is what I was working on last night with Agent Zero. And we can just see we have all these memories. So it's like solutions. I don't know what that is. Start free. I make a lot of sales copy and stuff, and I brainstorm a lot with Agent Zero for the AI Captains. So here's a main memory, which looks like it's more of like a knowledge-based file, but it keeps it in here. And look how in the memory we see retrieval tags. This is where the RAG, retrieval augmented generation comes in. The query, if I ever search for AI Captains PRD or school community structure or high ticket guarantees, avatar scalability, launch timeline, it will find this memory. So let's do that. And I'll copy that. I'll start a new chat here. And I know I'm zipping around, but I just want to show you how that works. I'll say, tell me about the AI Captain School community structure. And now you'll see the first thing it's doing is searching memories. Okay, it's searching memories. We can expand its work here. Do I have the model set right on this one? No, I don't. Whoops. Okay, let's try that again. We'll nudge it. Searching memories, generating, okay, thinking. The user is asking about AI Captain's school community structure, which is a well-documented topic in the knowledge base. So it already looked through the knowledge base, okay? I have access to two key memory sources, the PRD version 2.0 and the school optimizer agent system prompt, both rich in structural details. I need to synthesize this into a clear actionable overview that covers avatars, pricing tiers, engagement mechanics, et cetera, et cetera. So now it's saying, whoa, the school optimizer agent, its knowledge base provides the most operational insights. And so it's looking at everything it has and it knows in its memories that we've talked about this before. And I think this is actually the memory that we just looked at, right? Pretty close at least. And now here we go. We have an output, the AI Captain School Community Structure. And we got all this information about it. Now, that is an example of how the memory works. And that was like a combo of the memory in here. And if we go into the files, if we go to the A0 folder once again, and we go into Agents, let's go into Knowledge, then we'll go into agents. In knowledge, we have school, and now we have all kinds of stuff in our school knowledge base inside of Agent Zero, and that's stuff I've done in the past. But it's referencing that when I mention school. So it's like, this is my Agent Zero for the school community, right? Whenever I want to talk about school, it just knows everything about the school. And I could probably install some kind of MCP or some kind of way to feed the data from the school into the agent. Hey, tell me what people are talking. If this was a bigger school, there's only 50 people in it, but if there were 5,000 people in it, hey, what are people talking about right now? Let's create content about that. So you could synthesize information that way with your agent, your agent, which is pretty cool. So that is what it pulled from, from the knowledge for those memories. But then it also has this school optimizer agent. So the school optimizer agent has its own thing going on. It has templates, it has a system prompt. And so that's also where it pulled the information for this response. So all of that together is a good introduction to how Agent Zero works. You have agents, they have custom prompts, there's knowledge-based folders, and then the final bit that otherwise, that connects it all together is the behavior.md file.

So here in the A0 folder, we have this, where is it? That's weird. Okay, I can't find it. So I say, what is the location of your behavior.md file? And it will use its own little command line in its own little local container. Oh, and templates. Behavioral rules. So, oh, this is in American English. Well, maybe the new version changed that because in the old version, it was British English. Well, that says it doesn't exist. So that's not a good look, Agent Zero. So let's look at that real quick. It gave me that answer based on the system manual. It has its own manual for itself like made by Eon. It has that inside its files. So it actually references that when you ask questions about its own infrastructure. So it's saying, okay, that contradicts the system documentation. So now it's going to execute code, which is, hey, help me find. And it's not there. It looks in the readme. And then it did a search. Multiple files with behavior in their names, but the relevant one is this one. Okay, a0 prompts agent system behavior.md. That may or may not be the one I'm trying to use, but what we'll do here is we'll just click it and it'll download it. So then we'll open that up. Rules, no, it's not there either. Now I'm just going to say search for behavior.md. It really should have a behavior.md because I told it just like yesterday. Always use that school agent whenever there's something going on. Tell me your behavioral rules. Let's look at that instead because I can't find the file.

Okay, so this isn't exactly what I wanted to show you, but I did want to get to this anyway. Right here, we have the agent system behavior markdown file. So agents are really just like hierarchies of prompts. So you have Agent Zero that has this prompt, which we're going to read in a second. Then it will spawn an agent one that has this prompt, which is maybe just like Agent Zero needs code for what I'm asking as one task to accomplish what I'm asking. So it says, okay, agent one, you're going to make this code. So then Agent Zero has its big prompt. Then agent one just says, write this code. And then in order to write the code, agent one needs actually agent two to go do some research. And so agent two's prompt is just go do this research and then tell it to agent one. And then with that, agent one can write the code and then give the code to Agent Zero, which can then continue on with the request that me, that I, the user, made. So I hope that makes sense, because that's what we'll call the hierarchy of prompts. That's how agents work. It's a hierarchy of prompts. And so this is like the utmost prompt that Agent Zero uses. This is its behavioral rules. It's like a big system prompt. And we'll also see here, we have some of the stuff that I have added to. So favor Linux commands for simple tasks instead of Python. That is default. That's not me. Use subject lines for emails from this file. That's me. Validate every email against this. That's me. Write tweets using the templates. That's me. When the user says email blank to me or email me, send the email to jordan@aicaptains.ai with the Mailgun MCP. That's me. Generate AI Captain School images with this AI Captain's content agent and automatically email the images to Jordan. That's me. Okay, so all that's me and it hit a bunch too. So now let's keep scrolling and further down we see really like the bread and butter of any agent, but Agent Zero here. Avoid repetition and insert progress. Never assume success. Use tools to simplify tasks and achieve goals. Save useful information with the Memorize tool. Present results and verify with tools. Don't accept failure. Retry with high agency. I almost want to delete that because sometimes they keep retrying and going the wrong directions, right? But all of this is what gives, I would say, like the personality for what Agent Zero can do.

So here we see tool usage. And here, literally, like the first few, use the correct tool for each task. Never use code execution tool for coding tasks. Use Cloud Code MCP instead. That was what I added yesterday. Use Call Subordinate tool to delegate specialized tasks. That is built into Agent Zero's framework. Use A2A chat for communication with other agents. That's another built-in tool. But Mailgun here, Twitter MCP, Firecrawl, YouTube transcript, document, not document query, but those four, those are all MCPs I added. So as soon as I added the MCPs, then Agent Zero adds it to its behavior, its system behavior saying, hey, these are tools you have now. So when the user says, post this tweet, it automatically knows what tool it will use to post that tweet. Same thing with document query tool for document analysis. Use search engine for web searches. Use code execution tool for terminal commands, package installation, or non-coding tasks. And that is because I told it to only code with CloudCodeMCP to save me money and just to do a better job than the open source agent. So then it has a list of all its specialized agent profiles. Okay, when you need to call a subordinate agent, here's the hacker agent for cybersecurity and penetration testing. Here's the school optimizer for school community building for revenue maximization. Here's the researcher agent, research, data analysis, and reporting. Here's the developer, complex software development. Here's the AI Captain's content creator. Here's the other content creator. Those are my agent profiles. So as soon as I make a new agent, which I'll show you how to do in just a minute, it adds it to its behavior so it knows. It references its behavior at every message, basically. At the start of every conversation, it references what it knows in its behavior and memories so it can perform for you the best possible job that you need done. It keeps its memory. It keeps everything in there so it knows you. It knows its rules. It knows what you want. Here are some constraints it's got. It adds that. Note, these rules are the authoritative source for Agent Zero's behavior. It's in this file, agent system behavior. It's a template that references these rules. So this changed in the new version. That's why it took me a few minutes to find it. It used to just be behavior.md. Now it is within the agent zero prompts folder, the system behavior file. So that is referenced at the start of every conversation. So here it is in the response. It gives me everything it has in its rules, how it manages its memory, how it executes tasks, how it saves files, manages the paths, tools, agent profiles, constraints, etc.

So I showed you in the files here in the A0 folder, we have all these sub-agents. Let's make a new one now. So I'll say, let me think. We want to create a new sub-agent that can, I want to use, so inside my MCP servers, which I can open here, I have sequential thinking in here. Sequential thinking right here. And what I would like to do, and we can see here also, we can see all the different MCP servers that I have installed here. And what I'll say is we want to create a new sub-agent that can create plans for delegating other agents and completing complex sequential tasks. This agent should be called when these situations arise. This agent should always use sequential thinking MCP. Okay, real quick, I just want to invite you to join the AI Captain School. That's what this interruption is for. AI Captain School. We have workshops like this every week. We have different things going on. This is the Agent Zero workshop. Next week, we have a spec-driven development workshop. And there's a ton of cool courses too. These are sneak peeks. They're still in draft mode, but there's a lot you can learn here to build your business, whether you're building a directory and you need to web scrape, whether you want to learn the command line and learn how code actually works so you can build actual apps with AI. Whether you want to automate, whatever it is, we have resources for you to learn as well as a thriving community where you can ask questions and we talk about stuff. So if you're a solopreneur or an online builder, and that sounds interesting because AI agents are the future and you don't want to get left behind, then I'll see you there.

So now we're simply just telling Agent Zero we want to create a new sub-agent that can do this stuff for us, create plans, and it should be called when the situations arise for completing complex sequential tasks. And this agent should always use sequential thinking MCP. So I'll just send that off here. And one of my pro tips for Agent Zero is when you're really trying to build a solid infrastructure, you can get away with Ollama open source models here. But I would recommend using Anthropic or even OpenAI or even Grok. The proprietary models just seem to do a better job, especially the agentic models like Claude Sonnet 4.5. They just do a good job of understanding a little better what it is that these agents are supposed to do. So for example, this is what happened after I sent that message. Agent Zero created the new profile, new sequential planning agent profile. And then here are its thoughts. I need to do this. The agent must do this. It should be triggered automatically. I'll define a new profile. So it used the tool Call Subordinate. It actually didn't use that tool, but it's like under that, it falls under that tool. The profile is called sequential planner and it made a system message. You are a sequential task planner. Your role is to analyze complex tasks, break them into sequential subtasks and delegate each subtask to the most appropriate specialized agent. You must use sequential thinking MCP for all planning decisions. Never write code directly. Always delegate coding tasks to Cloud Code MCP. Follow all Agent Zero behavioral rules. Your output must be a clear step-by-step execution plan. Okay. So I created that prompt. Then Agent Zero called that subordinate agent, the profile sequential planner, and it gave it that message. You are a task planner, blah, blah, blah. And then that agent is now agent one. Okay. So we've gone from the Agent Zero, the top of the hierarchy to its first agent that has spawned agent one. I am the task planner agent tasked with breaking down complex tasks. I must use sequential thinking for all planning decisions. I cannot write code directly, blah, blah, blah. So So it knows. And it used sequential thinking here. I don't even know what it used it for, but it's ready. It's here and ready. Agent Zero responded. It's ready. Now that response is actually not going to me. It's going to Agent Zero. It's, what would you call it? The leader agent. So agent one is a subordinate agent. Agent Zero is the leader agent. I don't know the proper term for that. Ken, what's that called? Call it a manager agent. The manager agent. There we go. The manager agent, the ordinate agent, not subordinate. Okay. So now then Agent Zero says, okay, cool. The sequential planner agent, it was created and it's ready. It responded to me how it's supposed to, and now it's ready to go. So let's go back to the files. We'll go, we're here in the A0 folder agents, and we'll see. I don't see it. Now, good to know. So, I'll ask now, where is this new agent's profile files folder? I don't see it in A0 agents. Now, this right here is exactly why I recommend using Anthropic or OpenAI or XAI models when you're building out the infrastructure of your Agent Zero agents and workflows, because for some reason, the open source models just don't seem to remember everything after they examine all the documents. It knows it needs to put the agents here, but for some reason, it just didn't do it. And I've noticed that with these open source models, I get this problem more often. I have to remind it. You don't just make the profile. It spawned its own agent up here. And welcome, Alexandre. Oh, there's two Alexandres. Welcome. But basically, the Agent Zero here spawned its own new agent based on my prompt. I said, we want to create a new sub-agent. Now, in its defense, I didn't tell it to save it into the Agent Zero files as a new agent. I didn't tell it that. And so it literally just spawned the new agent, maybe because it thought I wanted to use it here and now. So that's probably why. But using Claude Sonnet, for example, normally Claude Sonnet is so ambitious. It just knows like, okay, we got to put this into the whole, we got to add this into your files because Claude Sonnet will just take it, everything to the next level. So that's why I would generally recommend doing that because then you have a little more leniency with your prompts.

So then I asked here, where's this new agent's profile folder, I don't see it. So in our thoughts, we see the profile sequential planner exists only as a named configuration in the agent's runtime context, not as a file. I need to clarify this distinction to the user and explain how it works. So let's see what it came out. It does not exist because it's created as a runtime profile. In Agent Zero, so now it's educating me on its own infrastructure. In Agent Zero, profiles like sequential planner are named context templates stored in memory, instantiated on demand via call subordinate with reset true. They're not file system based. Now, this is news to me, and it could be because of the new version of Agent Zero. It used to be everything was in these files. If we click the memory here, now we see in this fragment, we open it up, the sequential planner agent was created. And this is what it does. It analyzes complex tasks, breaks them into sequential subtasks, and delegates them to specialized agents. So theoretically, now, this should do the trick. If I say make a plan and delegate new agents, it always scans its memories in the beginning. So let's see what that looks like. And in fact, earlier here, it actually gave me some examples of how to, yeah, to invoke the agent. Whenever you request a task requiring coordinated agent delegation. So, uh, multi-phase research. Okay. Let's do some multi-phase. I just want to see what it does. I'm going to give it a very bad prompt here. I really just want to make sure it calls the agent from its memory. Let's see what happens. Let's do some multi-phase research on, on what? On Agent Zero perspectives, blog posts, et cetera. All right. So now let's see if it calls our agent. See, the first thing it's doing now is searching its memories, which once again, it has a memory dashboard here that it's searching through and you can give it a threshold. I don't even know what that threshold does. Maybe if it's for the search, but as we went over earlier and I'll talk about this just while this runs, we have a context window space that we give it in the setting. So itself, like the response, the thinking, the cleverness. The 30 remaining percent is used for the system prompt, the retrieval augmented generation, and the response. So for example, in these memories, where was the one we saw earlier? That was a good example. Was it in solutions? We see historical notes. There was a really good memory here. Yeah, main. So this is the main memory here, and we'll see the tags it uses for the retrieval augmented generation. Implementation tools, monitoring tools. Where are those tags? All right. Well, I guess I'm wrong about that. It's not doing that all the time. It's just doing it when it feels like it. But whatever one I referenced earlier in this video, it had a list of retrieval augmented generation, like the tags it would use to retrieve those memories. Okay. So here we go. It did all this research. Let's see if it used our, the real thing is, yeah, did it use, okay, the user wants multi-phase research. All right, it didn't call the agent. It did not do it. It just started calling the tools. It used the MCP to scrape the official zero, official Agent Zero website. Okay, so that didn't work. So now I go back here and I say, well, it doesn't work, motherfucker. You need to examine the file structure of the other sub-agent profiles in A0 agents. I believe that's what it is, A0 agents, yes. And create a new directory in the same format for this sequential planner agent. I'm 90% certain this is happening because I'm using an open source model and I'm not using Anthropic. But I spent $50 the other day building a pretty basic tool and I want to save money. So I'm using the open source models, but I'm pretty sure that's why it's happening. All right. It did what I said. You'll notice the user is insisting on a

File system based agent profile, which contradicts agent zero's dynamic profile system. So I'm going to have to learn more about that. I'm not sure what that's all about.

I do know that these sub-agent folders in the A zero agents folder work. I do know that for a fact. So that's what I want. So I created these folders. Let's see it now. Directory now mirrors the format with the right files. The system main.md file contains the required instructions for the agent. Great. This satisfies the user's request. Great.

You'll also notice agent zero isn't very sycophantic. It's not saying, oh, you're right. I'm sorry. Or like, hey, great idea. It's just doing the job. And it doesn't give me any banter. And that saves on token costs, of course, with APIs. But just in general, it's like kind of refreshing because when you're using proprietary models all the time, they can be a little annoying.

So now we'll go into the agent zero files and we see sequential planner. There it is. Boom. Prompts context. So now we'll go back to our agent zero research prompt. I'll copy that and we'll start a new conversation. I'll paste that in and let's see if it calls our agent. Oh, that's hilarious because this is what's cool about Agent Zero, everyone. It already did this research and it wrote a whole blog post about it in the AI captain style because it has all that jumbled up in its memory, in the things I've done, in the different agents and system prompts, and it just assumed that's what I wanted, which is hilarious, right? That's not what I wanted, but I just decided to do that.

Now, when you're used to Cloud Code or these other agents, you can get a little frustrated with this, but Agent Zero is supposed to be like this, where if you don't give it the exact correct prompts, it's just going to do it. And now what it is, is who knows, but it's just going to do it. And that gives you the opportunity to really refine it, because then in the future, it'll keep doing it. And it won't change because that's just how it works. And it's persistent memory here. We'll just keep it going. So if I were to say, no, clear your memory on this and do the multi-phase research again, let's see what it says.

I also want to mention that this is also why you might want to install multiple versions of Agent Zero. So here is the new installation we did since we started this call and I didn't get these set up right. But if actually it would work now, let's say, hi, what is your purpose? That's still not set up right. But you could have multiple agent zeros running in your Docker container. So here in Docker, I have two agent zeros working right now, the new one and my original one. So I could have one agent zero setup, be a template for my virtual assistant, right? Make the content, send it to me on Telegram so I can post it at these days. We have recurring tasks here, which I haven't gone into yet, but it can do things every day at a certain time for you. And it just keeps that in its memory. So I have this school optimizer do weekly research for me. It scrapes Reddit and updates pain points about the people that I think would do well in the AI Captain school. I always scrape their pain points from Reddit. And then it updates its school database in the knowledge base, not database. It updates the knowledge base. So here we're in the files in the A0 folder. We go to knowledge here and we see we have school and there's a whole database here. We have raw data from Reddit, from our school stories. And so I can keep that going. It also researches people running school communities. So I can learn like new techniques for monetization or engagement strategies, whatever. So this can be my whole business assistant. But then the other Agent Zero setup could be, I mean, who knows? It could be my cybersecurity assistant. Like, hey, I just vibe-coded this project. Hack it. Do everything you can to hack it. Can you get through it? And it's set up just for that. It has MCPs just for that. And that's what it's for. And then in its memories, it will remember. Like, this one is basically my AI captain's assistant at this point. Because it will always just remember what I'm talking about, what I'm building with AI captains. It's in its persistent memory until I press delete, right? Until I remove it. So that is something to keep in mind. It can be annoying or it can be super powerful or it can be somewhere in the middle. Hopefully not annoying at all, but really just productive. But this is the granular control and the attention to detail. You're able to give your agents with Agent Zero because it's open source. You can't do that with Cloud Code or any other proprietary solution. And so as Jan mentioned in the beginning of this call, when he introduced this, it's really important for there to be an open source alternative to powerful AI agents because in a lot of ways, those are very opaque boxes you can't see out of, you can't see into, you're just kind of hoping you throw the CloudMD and hope it sticks. But this is like literally its whole software that it runs in a Kali Linux environment in a Docker container. It has full control over its own little mini computer here. And it just will remember everything for you in an efficient way. And in the new version, there's projects. So there's going to be a new button here that says projects. So you can have different projects. And that might render irrelevant everything I just said about having multiple agent zeros installed.

Okay, so let's see what this did here. Did it call our agent? Let's go up. All right, I delivered the blog post, and that's when I said no. Initiating fresh research. Okay, I just went straight to the search engine. Okay, so now we go to a new chat. I say, if I ever want multi-phase research, add to your behavioral rules that you must call the sequential planner sub-agent. So it wasn't calling the agent like it was supposed to. Once again, I think it's because I'm using an open source model here in the background, not Anthropic, not CLOD.

So now it's updating its behavioral rules. The user has requested to update the rules automatically. I need to verify the profile exists. The existing solution memories show that the profile doesn't indeed exist. It's defined and functional. And since the sequential planner our profile is already established and documented, I can safely update the behavior rules. Okay. So now it uses the tool behavior adjustment to adjust its behavior. And it added that to the file that we went over earlier, the system behavioral rules. That's everything that it's programmed in by agent zero and then everything I've added to it. And now it will auto delegate to that agent. Okay, so now hopefully we've done our debugging here. Let's try it again. The only way it can go wrong now is that it added new memories from that. Yeah, gosh, it's not doing it. All right, let's play a game here. I'm going to use Anthropic now and let's see if it works. And my suggestion to Agent Zero is to make this easier, like to switch through models. So you don't have to like type it out every time. Okay. If I want multi-phase research, what is your behavior program to do? Okay. So it knows. I should clearly explain that I automatically delegate multi-phase research. So now it's passing that response on. So it knows. All right. So what's going on here? What else should we research? Ken, what else can we research? Give me an idea here. There it is. There it is. Okay. There you go. Spectre of in development. Let's see here. According. Okay. Finally. All right. Now, once again, we're using a proprietary model that excels in agentic tasks and Hey, suddenly it's working, right? The user is asking for multi-phase research on specter of development. According to my behavioral rules for any multi-phase research task, I should automatically delegate to the sequential planner sub agent. Bingo. We did it. Okay. All it takes is changing up the model or maybe changing up the topic. I don't know, but it's calling the agent now. So this is the fun part. Agent zero here, calling subordinate agent. Boom. Agent one now is the first subordinate agent and it creates a systematic research plan for the spec driven development study. All right. So it uses the sequential thinking MCP tool and it creates six distinct phases for this research. Core concepts and definitions, methodologies and frameworks, benefits and challenges, real world applications, best practices and tools, and comparative analysis. I need to create a systematic approach to gather information for each phase, ensuring through coverage and logical flow. So, Agent 1 then makes that plan here, and it uses the sequential thinking MCP task. It's planning phase 1, phase 2, phase 3, each one in a row. And it's got phase four, for example, search for spectrum and development case studies, API first companies examples, open API success stories, spectrum and development real world, target companies known for API first approaches, et cetera, et cetera. Planning phase five, planning phase six, and now it's executing phase one. So it made the plan and now it's executing. Now at that stage, I'm probably just going to pause this because I don't really want all this research right now. But really, if I were really serious about making these agents work, which I kind of am, is I don't want agent one necessarily to execute it. I wanted agent one to plan it, to make the plan.

So now we would add to its behavior. Let's see, how far did it go here? Yeah. Okay. So now I'm going to open a new chat and I'm going to say, create a new agent profile. And I'm just going to reference it this time. CA0 agents to make sure it matches the structure of other sub-agents. Create a new agent profile that does research using Apify MCP actors. It can, but here's the thing. We actually already have a researcher agent, I think. So let's not do that. Researcher, yeah. So let's say add to the researcher sub-agent to do social research using Apify MCP actors. Also add to your system behavior to always use the researcher agent when scraping the web to find information. So the reason I want that is because then agent one here won't simply be calling MCP tools to do the research. It will call agent two, which will be the researcher profile to do the research. And then ideally I get to the point where it can potentially, I don't need, I don't know if agent zero does multiple agents at the same time yet. It might, but Cloud Code can, for example, it'll put the agents in tandem. So sequential planner goes, as soon as the sequential planner makes the plan, then it says, okay, research your agent, you do that. Reddit scraper, you do that. Proposal writer, you do that with this stuff, and then we're going to update it after or whatever. And so all these different agents work in tandem. I think that's kind of what the goal for me would be in something like this. If we're kind of trying to do spec-driven research, I want agent one to then spawn agent two to do that research and then report back to agent one. And then eventually agent one will report back to agent zero with the final results of the research. And then agent zero will be like, yo, Jordan, here's the research. And all I have to, and we're seeing it here, all the delegations and everything. And I, of course, can minimize this stuff so it's cleaner. but like eventually we will never know what's going on in the background. I mean, we can know, we can set it all up, but if we continue using the proprietary agents, we'll never really know what's going on in the background. And you know, there's a whole conversations regarding privacy and stuff, but if we want to have full control, not to mention privacy, but like really know what every little thing is doing, then we can build something like that with agent zero. And this is a good way to really explore it. And so eventually, once I've built all these agents and I know how they work, eventually this kind of interface, I'll only be interfacing with agent zero. I won't know who it's calling, agent one, two, three, four. I don't care. Just get the job done, right? Just get the job done and use what you can. And so what originally today's call was going to get into, and I guess we're kind of getting into it, is eventually you won't need to be a human saying, make a new agent profile for this, please. Do this and then fine-tuning it like I'm doing and making sure. Eventually, Agent Zero will just make its own agents on the spot. Oh, you want multi-phase research? Okay, well, we're going to need a sequential planner. We're going to need a researcher to find the keywords on Google. We're going to need blank, whatever. And then it'll just program code, write out prompts for its own agents that it needs for this specific task. it'll delegate to those agents and it'll probably save those agents for later once again in the memory right so it knows that these agent profiles are available and then boom it'll get all this done for me so that's the future of where this is all headed and the reason i love agent zero is because it helps us understand all this from the ground floor and play with it we can literally i'm playing with this whole thing on the ground floor here so i hope that's clear i mean that's kind of the takeaway that I really want to give you with this intro to Agent Zero is like you have so much granular control over what's going on that you can really build some cool stuff for your own unique purposes. You don't have to find templates. I mean, eventually there will be templates, I think, for Agent Zero, which will be awesome. But you don't have to just find a tool or a SaaS out there that will do something for you because this will be so easy to program, especially once you get the hang of it. It is just a few back and forth and suddenly I have a whole string of agents doing things. We talked about the AI SEO bot recently and the AI SEO bot, SEO bot.ai, well, whatever it was called. It is a really cool product for generating a SEO content. And there's like 10 agents as part of that software that are strung along. So the first agent takes all your avatar information, your pain points. The second agent researches Google for that stuff. The third agent looks at the, you know, whatever Google search console for what's trending. Like it has all these agents going to road. Then one agent writes it. Then one agent edits it. Then another agent looks at it from an objective point of view and criticizes it. Then the next agent rewrites it, like agent after agent after agent. And you can set all this up here. Easy. So then take this to the next level. Imagine you build those agent strings. And this is the hot sauce, guys. This is what you came for. Imagine you take agent zero, ditch the interface, right? Ditch the interface after you set it up. Make a new interface, make a web UI that can take your string of agents, in this case, the spec-driven researcher, right? Create an interface for users that pay $10 a month or whatever it is. Imagine then you can use Agent Zero to create your SaaS, to create your service, slap on a different interface, and it'll still work this way. You just have to make sure it's good enough. You have to go through what we're doing here. You have to go through every step, make sure it's doing what you want, fix it, adjust the behavior, but then just slap on a different interface and you have a total software as a service. And Agent Zero is open source. So you could, you can do stuff like that. Like the license allows you to do stuff like that. So there is a lot of potential here. You know, it's an open source project. So obviously it's not as popular as whatever, Cloud Code. But I think as more and more people learn about it, and I'm hoping you watching this can get some inspiration from it as well. New stuff is going to come along. people are going to be doing cool stuff. This is one thing I introduced earlier. This is one project being built off of Agent Zero. It's a fork of something else, OpenLLM VTuber. But this is literally like to create a VTuber, an AI virtual streamer that can talk to you, like an anime girl or something. And you can stream it, you can talk with it, you can talk as it. And this fork of this project runs off of agent zero. So agent avatar, you can set up a whole AI VTuber with it. And this is only the beginning. There's going to be so much stuff like this where it's using agent zero and we build apps on it. And this one's open source, of course, but not all of them will be. And you know, you get creative and build stuff with it.

So I think here it created, it adjusted the behavior. I forget what we were doing here, but I think I've proven my point about why it's important to build these agents and how you can do it and all of that. So I'm going to show you just a few more things to wrap this up. This context button, this shows you like once we run out of tokens in the token limit, this shows you what's going to be passed on to the next conversation. So this is what it keeps here, which is going to be 30%. Remember, we put 0.7 on that slider. So 30% here is going to be this, and that's how long they keep this. So it keeps the behavioral rules here in the system prompt. So it looked at the files, and then it added it to the system prompt at the very beginning of the conversation. What to do with transcripts? The Agent Zero system manual. So I mentioned earlier, there's a total hierarchy of all the agents, the Agent Zero software and framework uses. Agent Zero itself, this is everything. This is what it runs off of. You're an autonomous JSON AI agent. You solve superior tasks using tools and subordinates. You follow behavioral rules instructions. And you just keep going down. And this is all the, you know, genius of the creators here of Jan, of the guys going in here. Because really all AI agents are, are hierarchies of prompts, containers of prompts. So agent zero is the biggest container. Agent one would be a little smaller. Agent two, right? And then maybe they would even be like daisy chained or they would, they would, you know, you would visualize it a different way because all the agents aren't being spawned at the same time, for example. But this is the agent zero prompt. Like this is the no bullshit. You follow all these instructions. And look how long this file is. I mean, apparently it's 51,000 tokens. Response handling. You might be part of a long chain of subordinates. Avoid slow and expensive rewriting subordinate responses. Instead, use this include path alias to include the response as is. It gives you the profiles. And then what happens? So it tells us all about the MCP servers, why you would use them, input schema. It's like got everything here. And you'll never see this system prompt with Cloud Code. I mean, you could jailbreak it using, what's that guy, the elder plinus. You could jailbreak some of those proprietary agents and maybe see this stuff, but they don't want you to see it. They don't want you to know what's going on in the background. Here you go. Agent Zero, change it yourself. Do whatever you want to it. Open source. So there it is. Secret placeholders here. Everything about how the tool works is added to the context in every conversation. Once again, memory, memory, memory. It's all about memory. Agent Zero wants to remember things for you. It wants to do a good job.

So I promised we'd finish going through all the settings and we're almost done. We got all our models. We started, yeah, we went on a rabbit hole when we left the memory dashboard. But then after the memory dashboard, you have options for voice transcription and speech synthesis, which happens all locally. So it will install one of these models, you can request permission so it can actually speak to you or I can speak to it. I'm not terribly impressed with the performance of this yet, but let's see if this works. Did this work? Is it ready? And this interface is a little clunky. I'm not sure what's going on when that button is green and it always takes a minute. So maybe it sent, maybe it didn't. But eventually this will be like the ultimate intention-driven development platform because imagine when it's smooth sailing for just talking to agent zero. Okay, do this. No, do that. Okay, do that. And then not to mention, I do think it can also speak back to you. Maybe not. Maybe yes. I don't know. Not terribly into that yet with it speaking back to me. I'd rather read. But I do like speaking to the machines because I can do that faster. Much, much, much faster. I can communicate myself. So that's the agent settings tab. We've kind of gone through the external services tab here. You put in your API keys for the different services you would use. You can put, I'm not going to show it here. Actually, I'll show it in this other one because there's nothing to docs. Here you can put secrets. So this is like a.end file. You put API keys, you put passwords, whatever. So Agent Zero can access them. If you want to check your email, then you get put your email password in there, right? You can get put authentication in here. So you have to log in. I don't have that set, but if I wanted the password protected, I can. I haven't done this yet either, but this might come in with you doing some kind of a SaaS, right? Providing external API endpoints. So you can actually use Agent Zero as a tool that API calls can access it. And then give it a CloudFair tunnel so I can actually access my Agent Zero that's running on my laptop through the web. I don't have to be home on my local network to access it. So you can create a tunnel here, which is pretty awesome. MCP servers, we all know about that by now. You can make Agent Zero an MCP server. So if I have all these agents here, I can tell Claude Code, hey, use Agent Zero when you need to do sequential planning multi-phase research, right? Because I know Agent Zero already has that agent string set up, and I know it's working well. Or N8N, right? I could have N8N call Agent Zero as an MCP. Pretty gnarly, right? And then agent to agent servers, I have never messed around with, but basically agent to agent servers, like agents can talk to each other. So I can't speak to that too much about what that actually is. I'm still at MCP level. Developer tab, what shell interface is it using? Probably not relevant to too many of us watching this video. The task schedule, we can talk about that in just a second. I mentioned it earlier. And then backup and restore, you can create a backup and this will help you when you're doing updates. So when you update, you basically have to pull the whole Docker container again, as we saw in the beginning of the video, and you're going to lose a lot of your files. So this will help with that. And also updates are going to get easier in future versions. So I'm not going to spend a lot of time in that.

So if we check this tasks tab, and this will be the last little bit we do here, we'll see that it's got tasks. And I'm having trouble with the tasks. I'll be completely transparent here. But let's talk about in an ideal world where everything works as it should. Like I mentioned earlier, I have a school optimizer, weekly research updates with Scrapes Reddit. And it gives me data on what people are talking about, people that run schools and like people in AI communities, like what are their pain points, right? And then here, the daily AI captains image generator, It will create an AI captain's image as the AI captain's content creator sub-agent. It'll follow the system prompt I have for generating image in the right synthwave, vaporwave style. And then it will send it to Telegram, to that ID. It will not email it. This is the daily random image delivery. So that's pretty neat, even though it's not working. I got it working in a manual conversation. So for example, I could start a new chat. I could say generate an image for the AI captain school. It's about agent zero meeting an AI captain for the first time. And it'll automatically generate that in the style that I would use for the brand on the school, which is, as many of you have already kind of seen it, I've kind of got this new vibe going. Actually, you haven't seen it. This is sneak peek of version two. So it's much more Captain Jackson, agentic centric. And even in the call today, I made this fun image. Like, yeah, my agent zero prompts like made this image, right? And it follows the style and it's just always got that aesthetic and it looks the same and I can stay consistent and professional, which is awesome. And you can do that anywhere. I was going to do that with N8N and I realized this would be so much easier. It was just a quick chat interface. I don't want to have to go to an end-to-end, do the whole workflow. The other thing is an end-to-end workflow doesn't debug itself, can't fix itself. It's like kind OF a hassle to really set up an end-to-end workflow. And so I thought, there's a framework that's really important that's introduced in the AI captains version two. And that is like Builderby. The Builderby framework. Like what's your standing operating procedure? When you need a new tool and you have vibe coding available to you, what are the parameters? What needs to be met in order for you to build it or be like, nope, I'm paying for it. And so I use that as an example of do I want to build a whole N8N workflow and deal with debugging it in a user interface? Or do I want to build a string of agents that can do some of these automations much easier and simpler? because I'm going to have to type in, generate an image, here's what I want in it. I'm going to have to type that in anywhere. So at least if it's here and it already has this whole workflow down to a T, which is what it's supposed to have here in this recurring task, then I just say it. It knows how to do it. It already knows what style to follow. Notice how I didn't mention any system prompt. It knows exactly, look, calling subordinate agent, delegating the AI Captain School Image generation agent. So agent one is that new agent that knows how to generate images for AI captains. And it puts it in this directory. Let's see, is it done? Oh, we got an error. No idea why. But it also knows to use Venice for image generation. So it has a Venice API key in my secret settings. Actually, I do think it blocks that out so I can show you. Oh, it's gone. That's why it's not working at all because my API keys disappeared. I have no idea how that happened. So that's why it's not working. But that's what it does. And it would generate an image and it would send it for me. Yeah, API key needed, right? So I'd have to put the API key image, the API key in there. So I'm going to do that. But this pretty much ends my introduction to Agent Zero. So there's a few of you in here now. Axe has been here the whole time, but everyone else kind of showed up a little bit later. Does anyone have any questions or comments or concerns or something they'd like to say about Agent Zero before we wrap this up completely? I really liked it. And I'm sorry I was late. I messed up on the time zones and the change of time. But the way that you're talking about using this, I find really compelling and I could see would be very useful for me. Yeah, definitely. Cool. Glad to hear that. Yeah, I think really what it's all about is getting creative. You know, a lot of this stuff can feel overwhelming because of the infinite potential, right? Like you can do anything. You got to start slow and one task at a time, one feature at a time to what you want to build. Like right now I'm satisfied with this and I'm not satisfied with the recurring task that's not working. So I got to fix that. But I didn't jump from zero to recurring task. I started with, okay, let's generate the system prompt. Let's make an agent that does these. Oh, let's make an agent that sends it to Venice and then emails me. Oh, wait, let's make an agent that actually can just send to a Telegram group through a bot. So I don't even have to check my email. I can just show up on Telegram, which is an interface I use all day with my team and with whatever. I'm just starting Telegram more and more and more to communicate with my AIs. Agent Zero can't send me notifications to my phone. Maybe it could if I really hacked it together, but Telegram's on my phone. It's on both my phones. It's on my laptop. So it's like when I have an agent saying, hey, I did this for you, or hey, I need feedback on something, then Telegram, I get it right away and I can respond. Agent Zero, it's there in the background on my computer, but I'm not like using it as a chat interface. So I really like the telegram option. And here we go. It's like able to just do this and it's generating four variants now. And it's using the wrong model. I don't know why it's using the wrong model. Everything got all, use it or lose it. That's what my motto is now with AI agent stuff. Use the high dream model. Generate just one variant. Yeah. So I don't know in full transparency. I don't know why it's using the wrong model. All of a sudden, I don't know why my API key disappeared. So there's little things like that, that come up with open source projects. That's just how it goes. And that's why there's communities. I invite everyone to check out the agent zero school. It's free. There's a lot of people in here. People are sharing their ideas and what they're doing. Agent Zero also has a Discord. Click here. And the Discord community has nothing going on, and I just checked it out, and I was like, oh, wow. There's some stuff. I found it good for when I just have a question of why something's not working. I actually get a better response in Discord than I do in Agent Zero. Sorry, in school. The school community seems to be slower. I don't know. Once again, it's an open-source community, so it's not like people are here all the time living here. But, you know, I posted this a few days ago. I was like whacked out. I'm like, I spent $50 to build something for two hours of Vibe coding using Anthropic API. And I'm like, how do I use Cloud Code? Because that's half of my Cloud Code subscription every month. I spent it in two hours. So I brought that and I actually got a really good response. And I think this is actually really relevant here. And the first response was like, hey, these prices are going to get to zero eventually, right? Or close to zero. But what this person is saying, Dylan, shout out to Dylan, integrate a cost and token efficiency throttle. So in the agent system prompt, agent zero, maybe even using the MCP or A2A model switch when a task does not need the high complexity reasoning of Claude. And you basically added a system prompt, like don't use this model if it's going to be expensive, right? And if you're clever enough, you can do this. So what I told them is, look, that's a great response. I don't get it. I'm going to run this through an LLM that understands agent zero, and I'm going to have it write the prompts for me. But that's a solution that you cannot implement with codex, with cloud code, with grot code. You can't implement a solution like that because they want you to spend money. They want you to be using their product more. So of course, they're not going to give you an an option to change their system prompt. So you're spending less money, right? Like that doesn't go against business. That's not good business practice. So here's a good idea of how I can make my costs more efficient with system prompts with agent zero.

All right. So this looks like it got done. This is a PNG image, man. It's doing a lot here. It's doing way more than I want it to. So something's definitely wrong. It's also supposed to use Cloud Code for code, and it's not. So quick question for you. Are you just running this locally, or are you using this in your hosting environment so that you can have the thing continually running? So if you have your recurring tasks, it doesn't effectively put the sleep because the process is no longer running. So first off, on macOS at least, or whatever, So you need to set in your computer that Docker Desktop starts as soon as you turn your computer on. So this is pretty easy in macOS. You go to your system preferences. You go to users and groups, and you click login items. And then you can click add, the plus button, to add Docker Desktop. So that way, whenever you start your computer, Docker Desktop, the app will start. But that's not everything. But beyond the Docker desktop app, you also want the container to start. So what you need to do, and I'm going to go back to school because a very kind soul saw my first Agent Zero video and said, hey, here's how you do it. Because I was saying, I have no idea how to do this. There it is. Docker run restart always image name. So I could even just type this in. Docker desktop, how to always run an image. So you get this. Docker run, flag D, flag restart, always name, container name. You just put the container name or image there, image name tag. Fairly simple. Once again, Google it or ask ChatGPT or Claude to help you get that right. Just copy your container ID or, yeah, copy the container ID, basically. and then when you run that, as soon as your computer starts up, it'll automatically open Docker Desktop and it'll automatically start running the container. So if my computer is on, Agent Zero is on. If my computer is off, Agent Zero is not on. That's how it works for me. You can always set this up in a remote environment. Set it up on a droplet, on an instance. Set it up somewhere else, you can even use Cloud Code to set it up, right? You can say, Cloud Code, you need to set up Docker on this server, and then inside Docker, follow these instructions to install Agent Zero. And then on Install Agent Zero, you could figure out, I don't know how the proxy, the tunneling would work to access it. I'm not that much of a system admin, but Cloud Code could probably help you figure it out. And then you could run it remotely, and then you could access it from anywhere and it can just do things, right? Just like I run N8N remotely because I want it to run through workflows when my computer might not be on, right? So in the same vein, if you wanted to do things while your computer's not on, you might want to do that. So I just don't do that. I like having Agent Zero running when I'm on the computer. So I know, okay, Agent Zero's on and like these tasks here, it's nothing I need done in my sleep. Like these tasks that I'm trying to set up, I don't need them done in my sleep. I need them done while I'm working. And a lot of these, there's no rush. And if there is a rush, I just say, hey, do this now. I need this task now. And I'll do it now. And then it's done. So yeah, I think there's a framework involved. I think I sent an email. I did send an email two days ago about this, is when you're automating things and you have different options, NNN, Cloud Code, Agent Zero, you want to make a personal framework for yourself. Like, okay, do I need something on? Do I need something running while I sleep or not? Right? What you need running in your sleep? Yeah. Put it on a remote NN in my case. It's in a remote NN. Do I need it while I'm working and I want to be more conversational about it? And do I want it to evolve more frequently? I'm going to do it on agent zero because I can conversationally upgrade agent zero. It can upgrade itself, right? Just in conversation. Now, if I want to upgrade my N8N workflow, it could take days. I could have to debug a lot of it. It could just be a hassle. And so it's like, all right, this is something that's really easy. Like I'll show you, I do this for every YouTube video. It's a really simple workflow, but it doesn't need to be upgraded all the time, right? Yeah. So I just upload my transcript in the chat for my new YouTube video, And then it sends it to this AI agent, which has a system prompt. You're an expert YouTube content strategist. Here's the transcript. And then I have it set up to use Anthropic right now. I think Claude's on it four, but I'll probably use 4.5 now. And then it converts it to Markdown and emails me. And then I copy and paste that into my YouTube video in the draft as it's uploading. And now I have it, right? So that's something that's perfect for me, right? I could have have agent zero trigger this. So I have a web hook. In fact, I mean, I don't want to start a new project and maybe not finish it here, but for shits and giggles, we'll copy this web hook and I can put it into agent zero. This is taking forever, by the way. I'm concerned about how much API credits it's using, but it says it generated the image and it did. It emailed me the image. Yeah, and the attachment didn't work. So I got work to do. I got work to do on this workflow that was working, but now it's not. So something changed. So for example, this is a good point. Sometimes when you're working on building an agent, it will fix itself in the conversation, but it won't fix its files. It won't fix whatever might need to be in here, in the knowledge base, or even just in the agents folder. So in this case, it was the AI captain's content agent, And it has scripts to use. It has mailgun send, telegram send, telegram upscale. It has all this stuff. But I did that in the same conversation and it didn't update its own prompt. It didn't update its own behavior. So it doesn't even know it has those tools in all cases because I didn't, at the end of the conversation, I didn't say, okay, now update your agent profile and system behavior. So that's something to really keep in mind. You can't assume that it knows to update like everything you're doing and everything you're working on because it doesn't know. It doesn't make those assumptions all the time, especially if you're not using the big proprietary models, Cloud Sonnet, OpenAI, whatever they're at now, Cloud with a crock, whatever you're using. If you're using open source models, it's especially probably not going to do that stuff. So I didn't do that when I set all this up. And now I have to go back and find that conversation wherever it is. AI Captain School here, maybe. Yeah, there's the system prompt somewhere around here. And then I have to go back through and I have to bring that back. So that's just an important thing to keep in mind. Just because you start a conversation building an agent doesn't mean the agent is going to update that new sub-agent with every little step. Cloud Code would remember that. That's built into its prompt hierarchy. I could add that in too. You know, maybe this is a good idea. If we update an agent's profile or system behavior in the middle of a conversation, well, actually, if we update how sub-agents should operate in the middle of a conversation, always update your system behavior and agent profile and relevant agent profile, sub-agent profile immediately. And now I won't have this problem again, right? So that's what you're building. You're building your own agent from the ground up. And so these things we take for granted with Claude Code, we have to do manually with Agent Zero. So I think beyond the technology aspect of this, there's something important about the human aspect of just building something from the ground up and realizing that a lot goes into this kind of thing. And I think it'll help you take more pride in what you're building as well. Just as if you were like, I mean, imagine you're coding an app. Yeah. You're coding it yourself. Like every component, you're coding yourself. Like, wow, you're going to take pride in that. Now it's just like every component, we just have to prompt. And in this case, we're not prompting, we're not vibe coding an app. We're vibe prompting an agent that can self-spawn its own sub-agents, which once again can get pretty trippy. All right. I've added a new operational rule, mid-conversation behavior updates. When you request changes, immediately update. I'm even going to say it may not always be when there is, does this need to be? Oh, that's it. Okay. When there is a user request. Oh, no, it should be fine. When you request changes to how I or any subagent should operate during a conversation, I will immediately use the behavior adjustment tool. Great. Perfect. Just what we need here. So maybe I can just find this. Man, this is getting messy. I need like folders for all the conversations, right? AI captains marketing, behavior update request. Let's just say update everything again. And then let me open up Telegram. Did I get the Telegram that has sent it to me on Telegram? Nope. Yeah, so that was a... What a bummer. I was so proud of it. I actually think... I don't even remember what my YouTube content is anymore. I'm pretty sure it was in a recent video. Pretty sure that was my last video was how I built that. So that's kind of funny, actually. Yeah, so now it's taking the conversation and it's just making sure everything is updated. So that's cool. Yeah, eventually we'll figure out a good framework for building with Agent Zero and I introduced this call. Jan, the founder of Agent Zero, started the call with us, by the way. So he said, hey, and he introduced the project and his inspiration for building this. And I told him how for eight, since January, 10 months now, 11 months, I've been trying to find a good AI agent framework. Like AI agents is what got me starting to make YouTube videos and learning about AI again because I was like, okay, agents, cool. But everyone I tried just couldn't get it working. They wouldn't work like how they're supposed to from GitHub, from whatever. But agent zero really just worked off the bat. And it's a really easy installation just with a Docker container. And so I want to just have one framework that I'll use forever, right? I don't want to be switching all the time. And so there's definitely a battle here, a competition out there in the wild, like what's the best thing to use. And so my framework for deciding like what's the best thing to use is how flexible is it? Like how much control do I have over where it goes? And with future upgrades, could it possibly fuck me over with what I've already done? That's kind of, I got to work on that framework a little bit, but that's how I'm perceiving it is like, will it fuck me over? And like, do I have all control? And so far with this, it seems so. And I was just introduced to like II Inc. recently by an Agent Zero guy, actually. And this, which I do not like this bit, powered by FoundationCoin, but we are building the best open intelligence for everyone. So apparently they're working on something similar. But but apparently there's something real with their agent. Okay. So they have a beta out of their agent. Let's see. Do they have a GitHub? There's a discord integrations, start building now. Oh, so this is just a platform. All right. Well, that proves my point. There's all kinds of services out there popping up to do this kind of thing. I don't want to use someone's platform. I want to be able to have full control, open source, host it locally. One thing we talked about earlier too, is you don't know when governments might change laws regarding artificial intelligence. So open AI has already made it pretty clear that they want everyone's biometric data. They want everything about you. They want to create a super walled garden where if you want to operate inside of it, you need to play by everyone's rules. And so you don't know, we have no idea what's going to happen in that realm and what might suddenly be shut off. Like maybe I can't use cloud code anymore for political reasons. Like, what do I do? Right? Do I get left behind? My business fails? Like, cause if your business isn't going to, if your business isn't using AI, it's going to die. Competitors are going to beat you in some way, shape or form soon. Right? You have to be using AI. So it's like, oh, you can't use AI unless you follow the rules of a dystopian government. Well, agent zero solves that problem. And you have a good enough computer to run a local open source model. You can run this locally open source. You can still have intelligence, agentic intelligence, and you can still survive. So to me, that's really a foundational driving factor of why I want to learn this stuff and make it work. Even though obviously there's glitches and stuff going wrong, I don't care because I know I can fix it because it's built to be fixed. If it's not robust enough is because I haven't made it robust enough. So there's definitely a mindset shift here coming from Claude Code, which is super robust because there's a company with fucking billions of dollars behind it, making sure it's robust enough. So commentary that I think is important to think about as we build. And as solopreneurs, why are we building? We want freedom, right? We want financial sovereignty. We want to have our own business. We want to be in charge of our hours. and we don't want to be beholden to anyone else for making our livelihood. Now, how long is that going to take? It takes time to get there, but we want that. And I mean, that to me is just the same line of thinking as why you want sovereignty over your AI technology is because without it, your business isn't going to have that engine behind it that other businesses will have. So to me, it goes hand in hand with solopreneurship. Open source is possible. So you can have as much control as possible. So if the rug gets pulled out from you tomorrow, you can stay alive and you can keep on thriving. In fact, it doesn't even, it hardly changes anything about your business, right? If the rug gets pulled out from you tomorrow. So interesting food for thought here. And with that, I'm wrapping this up. So there it is, folks. Thanks for watching. Note that the projects feature was just added to Agent Zero the day after this workshop was recorded. So that was missing, but honestly, you'll get the hang of it. And also I'll make another video about it and it's all good. So go check out Agent Zero. It's pretty awesome. Come join the AI Captain's school. I recommend it. Join our community. We like to build. And the Agent Zero school is cool too. So shout out to the Agent Zero squad and And shout out to you for wanting to build sovereign AI agents. Good luck and Godspeed.