Transcription
I've been building AI agents since 2022, and in the grand scheme of things, that's not a lot of time, but for AI agents, it certainly is. I was super early on building agents when Chat GBT was just getting popular. We didn't even have things like function calling or other competing LLMs like Claude or Gemini. And it's gotten to the point after a few years where I'm now very confident no matter the agent that I have to build. But it took a long time to get there and a lot of very hard lessons learned along the way.
And so what I am super excited to dive into right now with you is a clear and simple process that I would follow today if I had to learn how to build AI agents from scratch. And I can guarantee that if you follow this, you will learn how to build AI agents 10 times faster than me skipping all of the hard lessons. And so no matter your experience level, you can feel confident building agents for yourself, your business, or other people. So with that, let's dive right into it.
So, here is a bird's eye view of the path that I would follow right now if I had to learn how to build AI agents from scratch. And I'm breaking it up into 10 different phases. And don't worry, I'm being quite specific in each phase. So, it's not like there is a lot of time that you would have to spend to go through this process. In fact, even if you are completely new to AI agents, you could easily get through everything here in just a couple of months. And I will stay pretty high level here as we go through the phases, but I will call out videos on my channel that apply to what we're talking about if you want to dive deeper into that specific subject.
So with that, let's go ahead and dive into the first step, which is phase one, building out our foundation. And this is starting with the absolute basics like what is a large language model? Learning about the best ones out there right now like Claude or Gemini, GPT, even learning about local AI. These are large language models you can run on your own computer like Mistral or Llama. And then also learning AI agents versus traditional automations. It's a whole paradigm shift when we add this extra reasoning into our automations through AI agents. There's a lot more power but also unpredictability that comes with this. And so I have a video right here where I dive into that. It's really important to understand that difference. And then also we get to effective prompting. Learning how to talk to these large language models and agents. And I am specifically leaving out the phrase prompt engineering because I think that is overused and overhyped, but it still is really important for us to learn how to talk to these LLMs just like it's important for us to learn how to talk to other people. And then lastly, leveraging out-of-the-box tools. And so I want to talk about this right now, even though this is a path for building our own agents, because it's important for you to build an agent that's not just solving a problem that has already been solved. And I see this happen a lot where someone wants to build an AI agent. They want to automate something with an LLM, but there's an out-of-the-box solution that already does exactly what they need for like $5 a month. So it's like, why invest the time into building that agent? And so do your due diligence before you build an agent, making sure there isn't something already out there that solves your problem completely or at least for the most part. So we have tools like ChatGPT or Claude Desktop for your chat applications. Things like MCP servers, which is a way to give capabilities very easily out of the box to our AI agents. I have a video here where I talk about that. I'll just show you really quickly a couple of the AI tools that I use out of the box. None of these platforms are sponsoring this video. So, I just want to show you at least a few of the tools that I use pretty much on a daily basis. Starting with Claude Desktop. This is my go-to application just to quickly have conversations with an LLM. Bounce ideas off. And they have support for MCP servers and other tool integrations to make it very easy to connect to my other AI agents. And then along with it, I like using this platform called Aqua Voice. So, you can see in the bottom middle that I'm speaking right now. And then it can take what I said and turn it into text in any text box on my entire computer. Take a look at this. It is just so powerful. I love using my voice to quickly communicate, especially to LLMs. And then also, I love this platform called MEM. This is my place to organize all my notes and things like scripts for my YouTube videos. And they have a great API to make it very easy to integrate with my other agents, pretty much just like Claude Desktop. And then the last one that I'll show right now is Perplexity. This is a deep research tool that I'll use all the time. When there are new AI tech that comes out, new LLMs or frameworks that I want to research, I will use Perplexity for that. So, those are just a few of the tools that I love using. The last thing that I'll say here for phase one is that you always need to remember capabilities over tools. And what I mean by that is you want to focus on what you're able to accomplish instead of getting into the nitty-gritty details of mastering specific tools. And it's so important. This applies to all the other phases going forward when we dive into other things like no-code tools or AI coding assistance. You want to learn the high-level skills that will carry you far no matter the tool you end up using. You don't want to get caught in the weeds spending hours and hours and hours mastering very specific tools that might become irrelevant next month. And I've caught myself doing that all the time. This is really one of the biggest tips that I have for you in this entire video is capabilities over tools. And I have some very concrete examples of that going forward as well.
And so that will take us now into the next phase, which is building AI agents with no-code tools. It doesn't matter if you're brand new to automating or if you have a lot of coding experience. I would still highly recommend starting with these no-code tools. When you learn for the first time how to build AI agents, and there are a lot of great platforms out there for this, like N8N or Flowise, other ones like Voice Flow and Relevance AI. And I still use these tools all the time. Especially for prototyping for my AI agents because you can get your agents up and running so quickly with these tools. Even if you do have a lot of coding experience like me, it's worth using. And this is your time to learn how you can integrate tools with your agents so they can do things on your behalf and your services like Gmail and Slack and Outlook. Learning about retrieval augmented generation. This is a super important process to give external knowledge into your AI agents. It is essentially making them domain experts on your documents. And I cover RAG a lot on my channel. Here's a video specifically where I'm implementing RAG within N8N, which that is my favorite no-code agent builder. So definitely check that out if you're interested. Also, learning about basic agent memory is really important in this phase. We see it all the time with applications like ChatGPT and Claude Desktop where the agent is able to remember the conversation and previous messages that we sent. But it's important for you to understand how that works and how that data is stored in our database. And lastly, I'm a very hands-on learner. And honestly, I believe pretty firmly that everyone is. You don't get a lot out of a college course, for example, when you just have to read through a textbook and take a bunch of tests. And I've been there, done that. I always learn a lot more when I actually do something. And so, I'm going to be calling this out in a lot of the different phases. You want to learn by doing. And for phase two, that means building agents with no-code. So at least build one, hopefully many different agents with tools like N8N and get that experience, that comfortability before then moving on to our next step which is learning how to use AI coding assistance.
If you are not using AI to help you code your applications and automations, you are going to fall behind. Even Anthropic, the company that built Claude, 70% of the code that's written for their company is written by AI now. So this is not just hype. The future of creating software is definitely humans working alongside AI with these coding assistants. You have your AI IDEs like Cursor, Codeium, Replit. These are on your computer operating in your development environment. Then you also have front-end specific builders like Vercel AI SDK or Bolt. These will run in your browser. It's also very important to learn how to prompt AI coding assistants. And so this goes back to phase one, prompt engineering, building effective prompts, but also there's a lot more specific things to be learned for working with AI coding assistants specifically. And then also leveraging MCP servers. I already mentioned this once, but giving these extra tools to our AI coding assistants so they can do things like manage our database or search the web for up-to-date documentation. I've got a video right here where I cover my full process for using AI coding assistants, including things like prompting and MCP servers and setting up global rules as well, so you can dictate the behavior of your AI coding assistants. And then as far as learning by doing, again, going back to getting very hands-on here, build out some simple automations and agents with code. Download one of these AI IDEs like Cursor or Codeium. Build out some simple things. Just get comfortable with using these AI coding assistants before moving on to the next step here, which is using code to create our AI agents now instead of no-code tools.
And even if you don't have a lot of coding experience, I'd still highly recommend diving into this phase, especially because of what we had in the previous phase, using coding assistance to help in this process. And that's actually something I cover a lot in the AI Agent Mastery course that I have in the Dynamis community. So check that out if you're interested. More on that later. But yes, this phase is just so important because we can stick to no-code tools like N8N. It's very easy to use. It gets us started so quickly. I love using them for prototyping. And you can take powerful tools like N8N pretty far. But still, in the end, when you want the most flexibility, control, speed, and performance, generally I want to move to building my agents with code, specifically with Python because most of our AI agent frameworks like the ones you see below are using Python. Some JavaScript, mostly Python is the language for building AI agents. And I'm not a huge fan of "vibe coding." I think it's important for us to still understand what is being created through the AI coding assistance. And so I would recommend learning the basics of Python, still getting a grasp of programming, and then moving on to picking your AI agent framework. Like we have Pydantic AI and LangGraph. There are a couple of my favorites. We've got the OpenAI Agents SDK. We have AutoGen, CrewAI. There are so many good options out there. So explore, pick the one that's right for you based on simplicity and performance and the different criteria that you have. But also do not overthink it. Remember capabilities not tools. We don't want to get super ingrained in a specific framework. We want to learn the higher-level skills that will apply no matter the framework that we end up using because we don't know if maybe next month there's going to be some framework that comes out that just smashes all the rest of them. So focus on the higher-level capabilities. And we can do that by applying phases two and three. Like you learned how to integrate tools with your agents already, and short-term memory and retrieval augmented generation and using AI coding assistance. Apply all of that here. Learn by doing by now moving one of the agents that you prototyped with a tool like N8N into a coded version of your agent using a framework like Pydantic AI. So get that experience, get really comfortable coding agents and then we can move on to our next phase which is advanced architecture.
Taking our AI agents to the next level. This is where we get into multi-agent workflows and different architecture patterns for that. It's just so important for us for more complex systems to distribute responsibility between agents, just like it's powerful to distribute responsibility between people. There's an Anthropic article that I'll link to below that dives into that. Definitely worth checking out. And then we also have other components like memory systems for long-term memory, guardrails for input and output, fallback mechanisms. I have a link to a video right here where I dive into these different components. Guardrails is probably one of the most important ones though because we have input guardrails, making sure that we don't have garbage that's sent to our agent. And then we have output guardrails, making sure that there's no garbage that is sent to the end user. These kind of things are what is so important to make our AI agents actually reliable because like I said at the start, versus traditional automations, AI agents are more powerful but also a lot more unpredictable. And so we want to implement things like guardrails. And so learn by doing. Level up an agent that you coded in the last phase. Add in things like long-term memory or input guardrails. Maybe combine it with another agent and build out a multi-agent workflow. Just get that experience, that comfortability, leveling up your agents to make them more reliable and more effective.
And then once you get to that point, it is time to productionize our AI agents. That brings us to the next phase, which is deploying our AI agents. And the key for me for deploying my AI agents to production is to use a tool called Docker. This is where I can containerize my agents into an isolated environment that I can then deploy anywhere in the cloud, start getting users using my agent. There are a lot of cloud platforms available. So just do your research, find the one that's right for you. There are easy-to-use ones like DigitalOcean and Hostinger. If you're an enterprise shop, you'd probably want to use something more like Amazon Web Services or the Google Cloud Platform. If you need a GPU instance to run larger local large language models, then you could use something like Vast.ai or RunPod. And I show what this process looks like in this video right here where I containerize my agent and then I deploy it to a platform called Render. And for you, it's going to look very similar pretty much no matter which cloud platform you end up using. And part of that is in thanks to using Docker to create a very good standard for deploying agents. And so learn by doing by taking this agent that you've built throughout these phases already and deploying it to the cloud. Get comfortable with one of these platforms. Learn what it takes to build your agent out. Deploy it, scale it, get it into an environment for other people to use it.
But once your agent is running in the wild, it is very important to move on to the next phase, which is all about monitoring our agents. This is oftentimes called agent observability. It's our way to peer into our AI agent, which is especially important in production to see the requests coming in from our users, the decisions that our agent is making around tool calling, and then also the output that's given to our end users. And this is so crucial for us to detect any errors with our agents or just see opportunities to make our AI agent better. And we'll get into that in another phase, but being able to continuously improve our agents is so crucial. And there are a lot of good options for agent observability like Langfuse, Helicone, Langsmith, Logfire, Improv. I have this video that I'll link to right here where I dive into using Langfuse. It's not that hard to set up agent observability, but it'll take your AI agents very, very far. And like I call it right here, it is 100% necessary for production.
Speaking of continuously improving our agents, this brings us on to phase 8, which is agent evaluation. There is a very interesting statistic that I've heard quite a few times recently, actually. When you are building AI agents, it's only 25% of the actual coding or automating. And then the rest of the 75% is evaluating your AI agent and making adjustments to improve your agent based on the evaluation. And this is so important and also different from testing. So when we are testing our agent or just our code in general, we are testing for general code correctness. Making sure that our application isn't going to completely crash or the end user would not get a result at all. But evaluating is different. It's all about agent behavior correctness. And so if an evaluation produces a negative result, it doesn't mean that there's an error or an application crashed, it just means that our agent is not giving an acceptable answer, whatever that might look like. And there are a lot of different ways to evaluate agents. We have the whole idea of LLM as a judge, like using another large language model to evaluate the output of our agent. We have task completion testing. This is more around tool calling, like did the agent invoke the tools necessary based on this request to do something like update a task in Asana or draft an email in Gmail, whatever that might be. And then plain and simple, we also have human evaluation, having people like maybe these are our end users that fill out a survey after they use our platform and just trying out our agents, seeing if the results are acceptable to them. And evaluating your agents is really what's going to take you far. All of the previous phases are all about building a reliable agent, building one that's effective, that does the job, but then evaluating is where we can make sure that all of that is actually giving us the results that we need for the system that we're trying to build or automate.
And that brings us to the next phase, which is mastering AI with others. One of the biggest regrets that I have when I first started learning AI agents was not doing it with other people. I was learning pretty much alone and that did change eventually, but it would have sped up my process for feeling confident building AI agents so much if I did it with other people because mastering AI agents is not easy. There's so much power that's given to us with these LLMs. But with great power comes great responsibility. There's all that unpredictability with agents that we have to deal with. That's why it's so important to find a community and learn and collaborate with other people. And this is just the perfect time for me to talk about my community, Dynamis AI Mastery. I've been pouring my heart and soul into Dynamis over the past month, making it the perfect place for you to master AI with other people and collaborate as well. And you know that I pour my heart and soul into YouTube as well. So, I'm doing the exact same thing, if not more, with Dynamis. And I have the AI Agent Mastery course, which is where I go into my full process in detail for building AI agents from the planning phase to prototyping with N8N to building the agent in Python, creating a full front-end application and deploying it, my full process. So definitely check out dynamis.ai if that sounds interesting to you. And I already have hundreds of other early AI adopters in the community ready to collaborate with you. Daily events like office hours and workshops and more. AI agent templates and resources that are immediately available to you once you join the community and daily support and idea sharing. So, so much to offer. Definitely check out dynamis.ai.
And with that, that brings us to our last phase, which is leveraging your new skill set. Honestly, the thing I love the most about AI agents is just how versatile of opportunities you have once you get really good at building them. You can automate things in your own business or your own life. You can create and sell AI agent templates or frameworks. And that's a really powerful thing. By the way, creating systems around your AI agents, packaging up those and selling those is so lucrative. You can start an AI automation agency, which is kind of related to the thing above, but maybe making more bespoke solutions. You can build a SaaS product powered by AI agents. You can become an AI consultant or advisor for different companies. You can join or start an AI-focused company. Even at the enterprise level, this experience will just take you so far. You can teach others through content creation like I'm doing all the time. Contribute to open-source AI agent projects. There are so many different ways that you can leverage this new expertise. It takes you so far. And it doesn't matter the industry or the niche. Everyone needs AI agents. And so if you go through this full process, you get to the point where you're confident building AI agents, you can do anything. Like the world becomes your oyster. And that's why I'm just so excited every single day about AI agents. And I'm just so motivated to continue to put out things on my YouTube channel here, starting up Dynamis and building that out. Everything that I'm doing, I'm just so motivated because AI agents really is the future. That's why it's important to get on this now.
So, there you have it. That's my full process for learning AI agents if I had to start over today. I hope you found this really helpful. No matter if you're a beginner or if you're already partially through the path and I just illuminated the next couple of steps for you. And if you think there's anything that I'm missing in this path, please let me know in the comments. I love to keep working on this because the most important thing for me is that the hard lessons that I had to learn and wasting my time learning things that weren't important, I want to make sure that you don't have to deal with that as well. So, I'm just trying to pass this on to you. So, if you appreciated this content, you're looking forward to more things AI and AI agents, I would really appreciate a like and a subscribe.