How I’d Learn AI Agent Development in 2024 (if I had to start over)

AI agents are taken the World by storm in 2024 but AI agents cannot build themselves at least not yet even Tech giants like Google are recognizing the potential here with their own agentic platforms and it's not just Google a wave of stas products are now ditching traditional chatbots in favor of more advanced agent based systems and although some of these platforms haven't changed a thing beyond the copy on their lending Pages it's evident that the role of an AI agent developer will be one of the most sought after jobs of 2024 However unfortunately because it's

still so early no one really understands what exactly to look for leave alone how to become one so in this video I will finally demystify who an AI agent developer is what skills you'll need to become one and most importantly provide you with a complete road map that you can follow by yourself right now let's Dive In all right before we jump in I'm sure you might be asking who am I and why do you have to listen to me and what I have to say about AI agents so allow me to introduce myself my

name is arson and I'm currently running one of the most popular AI agent Frameworks on GitHub that allows you to create agencies with AI called agencies form but that's not all I'm also running my own AI agency where we recently launched the first of its kind agents as a service subscription so unlike other influen answers I'm not making my living selling you ads I actually make all of my income delivering real world AI Solutions in fact just in the past month I've hired two AI agent developers myself so hopefully this now gives you an idea

that I've got a pretty good understanding of what to look for okay with that out of the way allow me to break down what AI agents are I've previously talked about this on my channel so if you want to dive into more details please make sure to watch this video after here I'm just going to give you a brief overview because understanding how to think about AI agents will significantly impact your approach to building them from what I've heard so far typically most of AI influencers either Define them as automations or as your own employees

in my opinion both of these definitions are extremely misleading on the one hand AI agents have a lot more autonomy than automations AI automations run on your back end no matter what even if the whole world Burns RS down EI automation is just going to keep running like nothing happened on the other hand eii agents have less autonomy than your own employees because your employees can defy you you can provide them with as many instructions as you want but they can still consider things on their own and do them in their own way so the

closest definition to AI agents that I've ever heard was made by Andre karpathy who said that AI agents are essentially operating systems yeahi agents fall in between employees and out automations they have Ram which is memory they have access to tools and they can use other llms and reflect upon their own actions why am I telling you this well because often times I've noticed that some of you guys are trying to make your agents perform literally everything by themselves this in my opinion is wrong you shouldn't just tell your agent what you want and then

hope it's going to do it all for you your agent should only route the requests from the user to the tools all of the heavy lifting must be performed by the tools not by your agents think of it as your own OS like Windows or Mac where we have different apps for specific purposes we don't even consider how these apps function we simply control them and the same should apply to your agents you can't just tell them what you want and then sit back and wait until you become a millionaire unfortunately this does not work

out yet and at this point the most significant challenge that prevents Mass adoption of EI agents is their real ability currently to make agents reliable you need to carefully fine-tune them on a specific business process and this is exactly where AI agent developers come in as an AI agent developer Your Role is to ensure that your agents have access to all the necessary resources tools and knowledge to perform their designated tasks essentially what it entails is first conducting thorough research about the business and determining which of those resources the agent will need then you need

to create the tools as as I said tools are the most important while building your agency you have to constantly iterate in it gather feedback from the stakeholders and adjust it in the loop until it performs consistently from time to time and finally you must actually deploy your agents which is a step that a lot of people Skip and then integrate them into your client's business you have to make sure that these agents work seamlessly alongside their own employees now what skills required to become an AI agent developer well as for the soft skills definitely

communication and eagerness to learn communication because you'll often find yourself Gathering requirements from the stakeholders and setting clear expectations which is not an easy task and eagerness to learn because we're still in a very early stages with advancements in AI emerging all the time so you have to be able to adapt very quickly and as for the hard skills primarily it will be light backend development combined with AI so now can anyone become an AI agent developer in 2024 in my opinion yes anyone can do this because imagine if you were trying to excel in

a subject like math and you'd be up against people who' spend their entire lives mastering it however when it comes to AI agent development you'll only be competing with people who have at least a few months of experience so in my opinion the timing is just right even if you're a beginner now I'm sure that you guys might be wondering if coding will be necessary or if some no code platforms will suffice in the future well the short answer is yes coding will be necessary here's why first of all the most powerful tools that your

agent will use will always require some coding experience sure you can connect your agents to off the shelf's apis but typically in real world scenarios we always create those proxy end points even if we connect to Off the Shelf API this gives us a lot more control over the behavior of the tool for example we can process the input data and output data as we see fit moreover a lot of those tools do require multiple apis together or some custom logic inside secondly open source will always lead the way not in a sense of developing

the most powerful models but in a sense of actually using them in Creative projects SAS platforms often require thorough testing and a lot more time to release new features open source reers like my agency's War framework can adapt to new advancements in AI much faster and lastly man of my clients are extremely concerned about data privacy for them SAS platforms are not even an option they want to deploy agents on their own existing infrastructures so while SAS platforms will definitely evolve over time and they might satisfy some of our client needs in the future I

believe that learning to code will still be a non-negotiable moreover with the Advent of AI coding tools coding is literally becoming Ain to using another SAS platform because you can already code using only n natural language so the learning curve is nowadays much smoother and there is no longer a need to learn any complex data structures or algorithms you just have to know them on a high level so you know what questions to ask the rest AI can already do for you my road map factors in all of those AI coding tools so let's now

take the first step which is to find a good project yes guys I can't emphasize this enough you can start without a project if you don't use it you lose it everything that you learn without a project you will simply not retain on the other hand having a real world project from the start will give you your own lens through which you can interpret all of the following information and the best way to find a good project is to think about your own interests and how they could intersect with AI your Niche is essentially where

AI combines with your talents all of the obvious applications for L language models have already been capitalized upon if you were to create another EA agent for research for example you'd be up against tens of Open Source and Commercial projects providing this exact service however the real opportunity lies in merging AI with your own unique interests this is how you can find those untapped use cases that make Captivate other people just like you for example I used to be into EDM so I could create an AI agent that would send me the latest hits create

media files with a custom coded tool generate certain sounds with 11 laps API or even entire song when a company like yudo releases their own API so start by scoping an initial agentic project include how many agents you want to create what each agent will do and what tools they will use additionally it's helpful to determine the overall goals and the desired outcomes of the system if you want to download my notion template for scoping an agentic project that we use in our agency almost every day please feel free to use the link below it's

going to ask you for your email I'll explain why at the end of this video once your project is fully scoped you're ready to move to the next step which is the environment setup your development environment is where you'll be spending most of your time so it's important to nail it down from the very start first you'll need an IDE this is the software you'll use to write and test your code there are many idees to choose from but some of the most popular ones are visual studio code jet brains and X code I personally

prefer jet brains but it's not free although you can get it for free if you're currently a student next you'll need to install python python is the most commonly used language for AI development literally the entire AI industry runs on python it's also a great language for beginners because it's extremely easy to read just follow all the installation steps on their main website and if you run into any issues just send them to chpt please don't send them into comments after that you'll need to figure out package management this is how you'll install and manage

the various libraries and tools you'll use in your project messing up package management is the single most common regret for all python developers remember that you should never install packages globally for a specific project and always create separate virtual environments I recommend using a tool like K however python also has a built-in package manager called vnf finally you'll need to find your AI development tools as I said these are crucial in 2024 honestly I personally barely even C code myself I just check code provided by AI some of the most popular ones are GitHub copilot

cursor chpt Amazon code Whisperer and even your own agent like David I recommend cursor at the start which is an AI first IDE Fork from vs code it's one of those rare tools that strike the right balance between coding and using AI they are also backed by open AI once your environment is ready to go it's time to start coding so the next step is to learn the basics at this stage I do not recommend diving into too much detail you just need to learn programming at a relatively high level there are many resources available

online to Learn Python quickly you can find tutorials on YouTube or you can even read books like learning python which is what I did in my early days however what you can do now is also create your own agent or a custom GPT upload this entire book all of those YouTube videos and use it to assist you with your learning process another great way is to clone a repository on GitHub that you really admire and go through it with AI ask questions about specific lines of code that you don't understand then try to modify this

repo or add some new features feel free to try this out with my own framework and submit a PR if you create something cool next you'll need to learn git git is a version control system that lets you track code changes and collaborate with others this is important because sometimes as an agent developer you might have to collaborate with your client development team again there are many resources available to learn git you can find tutorials on YouTube or you can use tools like GitHub copilot CLI a new addition to GitHub copilot that explains or even

creates any terminal commands for you fireship also has a great course on git that I personally recommend once you've got the handle on the basics the next step is to learn the modern AI development text tag the first skill you absolutely need to master is an AI agent developer after you learn the basics is the use of llm apis currently the leading players in the field are open AI anthropic and Google Gemini in my opinion open AI is at least a few months ahead of everyone although they were extremely slow with releases recently so I

definitely recommend starting with open AI also in my opinion although the gap between open stores and commercial models will shrink over time commercial will always be ahead feel free to disagree with me in the comments so explore open a API and try to create an agent from scratch don't use any Frameworks or libraries equip it with tools using only custom code and try to fine-tune it on a simple task this will give you a much better understanding of how e agents work internally however don't stop on just commercial model apis sometimes there is a need

to deploy or fine-tune an open source model because of data privacy I am also currently working on adding a new way to use my framework with open source models and if this proves successful I will let you know so it's important to know how to reuse existing AI models in 2024 but I would not recommend learning to train your own AI from scratch as I said in one of my previous videos AI is just becoming way too complex secretive and expensive there is absolutely no way you can train any breakthrough model by yourself instead try

fine-tuning or deploying existing models on a platform like huging phas they offer convenient deployment options and a convenient way to fine-tune those models so I've actually released videos on how to do both on my channel as well and finally you'll need to learn function calling function calling is a feature that allows llms to call external functions this is how your agents interact with the outer world and this is why it's the most important part of the entire system in our agency we don't measure results of our systems by analyzing the outputs of the messages instead

we measure our results by analyzing the actions that those agents take I recommend exploring the instructor Library which is a project developed by Jason Leo that conveniently combines penic with function calls essentially it allows you to validate all the inputs provided by your agents before executing any logic inside those function calls so my framework depends on instructor for all custom tools and Json also has a detailed documentation with many great examples after that the next step is to learn multi-agent Frameworks the most popular ones are currently autogen crew eii and my own framework agency form

I've previously discussed all of them on my channel and I also released many tutorials on how to use my own framework so I won't go into too much details here if you prefer another framework there are also plenty of tutorials on YouTube essentially all these Frameworks do is just handle some of the underlying details for running your agent-based systems such as multi-agent communication function execution and State Management there is an intricate balance however between between having control over your systems and being able to get started fast my framework is definitely heavily skewed towards control and

customization rather than the ease of use it provides you with full access over your agents without any hardcoded prompts or any hardcoded agents even like in other Frameworks so other Frameworks might definitely be easier to get started with because you don't have to create all of those Proms and agents yourself but they might be too restrictive if you're building anything more complex than a trip planner now there is one crucial step that a lot of people skip which is actually deploying your agents in production this is why next you do have to learn some light

backend API development you can start by deploying a simple model via API there are many ways to do this for example you can use Firebase functions AWS Lambda or even create your own server with flask or fast API I recommend using serverless because the number of requests your agent receives can vary significantly from time to time therefore you need a scalable solution to learn quickly the best way is to chat with chat GPT it knows surprisingly well all of those infrastructure providers like Google Cloud AWS and Azure Google Cloud also has recently introduced their own

assistant powered by Gemini which can be used directly from their console you can simply ask how to create a serverless function and it will provide you with an exact step-by-step process to follow and if you encounter any issues of course you can send them back and it will help you to resolve it once you've deployed your agent as an API try integrating it into a channel like slack or Discord or you could even go farther and try creating your own chat user interface using a platform like verel VZ front-end development could be helpful sometimes however

it's not a requirement once you've shared your project with the world the final step is to monetize it making money is hands down the best way to learn because you also get paid for it and to do this effectively you must productize your service you must reuse as much of the previous code as possible you should create templates for each step of the process like deploying agents creating tools or even creating agent templates sometimes that could be helpful although I do not recommend using it all the time after that you need to find your first

client there are three means you can find your first client number one is freelance platforms this is the easiest and the fastest way to find a client without making significant time and money Investments I think AI agent development fits quite well with freelancing because the projects are relatively shortterm besides There is almost no competition for a agent deps on these platforms yet this is how I personally got started I used to work as an eii freelancer on upw work which gave me a ton of experience because I was working with many different companies on various

projects across all modalities like audio Vision NLP and so on my tip for you here is to create your own project offers and let your clients find you instead of sending proposals number two cold Outreach cold Outreach involves sending code emails and DMS to companies offering your service this method requires a lot more effort but it can also lead to higher paying clients I don't have much experience doing cold Outreach but my tip for you is to always schedule a call instead of trying to sell your solution via emails and finally you can find a

full-time job you'd be surprised how many companies are actually looking for a agent developers right now and the best way to find such a company is to look for startups that recently received funding our team is also currently on the lookout for agent developers so if you have already completed all of the previous steps you can reach out to us on our Discord remember the goal is not just to make money but to learn and grow as you go and to wrap this up I just wanted to share a bonus tip with you which is

to build your own community building a community requires a substantial time investment however if you have a community and if you understand the needs and challenges that your community is currently facing you can tailor your services specifically for them this will not only help you stand out but will also provide you with extremely valuable feedback so as I said I personally started as as a freelancer on upwork yet about a year ago I decided to start this YouTube channel I thought that maybe by this time I would get my first client from YouTube however believe

it or not my first client found me on my second video when I had only about 100 subscribers or so so I definitely recommend considering platforms like Twitter Youtube Youtube shorts medium or even your own open source project with a dedicated Discord server it will take some time to grow your audience but if you stick the investment will totally be worth it so yeah that's it for this video guys if you want to grab this whole road map and my AI agent scoping template for notion make sure to use the link below it's going to

ask you for your email because I'm currently working on a complete course that will include a lot more information that I couldn't fit into a single video here we are actually one of the very few companies in the world that productized AI agent development at scale so we have a lot more to share however before we do that I want to make sure that this course is as good as it can be so it will probably take me a couple months or so to make it thank you guys for watching and don't forget to subscribe