Tutorials · Tutorial

Build an AI Multi-Agent Team (Step by Step)

Viktoria Semaan · March 25, 2025

Transform your job search with an AI Multi-Agent team! In this video, you'll learn step-by-step how to build your own powerful AI agent crew, designed specifically to revolutionize job hunting and interview preparation.

Chapters

  • 00:00Why AI Agents Will Transform Work
  • 01:27What We'll Build Today
  • 02:36Prerequisites
  • 05:00Technical Walkthrough
  • 07:40Live Demo: AI Job-Hunting Crew
  • 10:50Cost Calculation and Considerations

Links and resources

Transcript

Generated from the video's captions and lightly edited.

If you want to learn how AI agents work and how to build your own team step by step, then this video is for you. You really don't want to miss on this because it's changing the way we work. Recently, NVIDIA CEO made it clear when he said IT department of every company is going to be the HR department of AI agents in the future. And he's not alone. Other executives are confirming that AI agents are taking on more and more office work. So today I'm going to show you exactly how to get started with AI agents and we're going to build a crew of AI agents that work together like a real team to help us to secure new job opportunities and prepare for interviews.

By the end of this video, you will understand how AI agents work and how to use them to solve complex tasks. It's been a tough job marketing tech lately with all of this layoffs happening. So I thought why not to build something that helps you to develop highly sought after technical skills while also solving a real challenge finding a job and preparing for interviews. So the concept we are going to learn today isn't just about job hunting. You can easily apply the multi- agent framework to other use cases and I actually created a GitHub repo where you can find examples for investment analysis and I will be adding there more use cases.

So find the link in the descriptions. Today we are creating what I call a job hunting crew. Here's how it works. You give our AI team a job posting you're interested in. Share your resume, maybe your GitHub profile or a personal website if you have one. Then our team of four AI agents gets to work. This agents use different tools to process all the information. They can read your resume files, do online searches, and share information with each other. At the end, they create a perfectly tailored resume that matches with key qualifications and highlights your relevant experience.

plus a customized interview guide with questions and answers based on the company you're interested in and your background. So, let's meet our AI crew. We have the job researcher who analyzes job listings and gathers insights. The personal profiler who structures your professional details. The resume strategist who aligns your resume with job requirements. and the interview coach who prepares you with a tailored guidance. If you've never built agents before, no worries, I will explain all the tools we will be using today. The only cost involved is running cost to large language models which can range from 1 cent to 20 cents depending on the model.

I will be running all my code locally using Visual Studio Code with a Jupiter extension installed. I also enabled Q developer extension which helps in case if I need to troubleshoot code directly in the ID. If you've never used any coding environment before, you can also start with Sage Maker Studio or Google Collab which run directly in your browser. No setup required. Now let's configure the prerequisites. We will use the Crew AI open-source framework for defining and running agents. To give our agents the ability to search online, we will use Serper, which allows them to find websites just like Google search.

You cannot sign up for a free API key. They provide credits. And since I started experimenting, I've only used very few of them as you can see on the dashboard. Then open the sample file and update your API key there. Our legends will be powered by large language models. To access this models, we will use Amazon Bedrock. I prefer Bedrock because it lets me experiment with different models using a single API for example entropic llama stability for image generations. To get started with Bedrock, you will need an AWS account. If you don't have one, create it first.

Once set up, go to bedrock and select a region. I personally always default to US West 2 because there are most of the models are available in that regions. But even if you are outside of US, you can still select US region. But you can also check regions that are closer to you and if models are available there. After selecting the region, go to modify access and request access to all models. Requesting access doesn't means that you will be charged. You will only pay when you actually call an API to use them.

Then install the AWS CLI and configure access to your AWS account. Check the documentation for guidance. I will provide the link in descriptions. So now we are ready to start building. First we going to set up tools. So tools provide us a way to extract for example information from documents like resume. You can also extend it to additional files such as different references about you. you also need to upload your resume and update the file name. for example here it's just called sample resume. Then select model to power agents. You can experiment with different models.

All you have to do is just change model ID and paste it here in the code. First tool is going to utilize retrieval augmented generation or rack. So documents will be chunked into text that then converted into vectors and stored in a vector database for ultraast retrieval instead of searching the full document every time. Then we're going to add a web search and web scraping tools and file loading capabilities. You can check Crew AI documentation for additional built-in tools or even create custom tool yourself. Now let's load a sample resume.

Once this is done, we can start outlining agents. So for each agent, we're going to define a role, a goal to achieve, tools that it can access, and a backstory, which helps guide the LLM for reasoning. We will define this for four agents, but feel free to modify goals and backstories to better feed your needs. Now, we can define tasks. Curly brackets in the descriptions means that information will be provided later. The research tasks takes a job posting and extracts and ancestral skills and requirement. The profile tasks scans personal websites and GitHub profile.

The resume strategist creates the tailored resume and uses input from previous tasks. And the interview coach generates interview questions based on the resume and job descriptions. Now let's assemble them to crew. We can assign a different LLM for a manager. Earlier we used haiku a smaller cheaper model for agents. For manager we can use codes on it. Finally we define the crew which includes agents tasks debug settings. For example here we can say verbose mode. And here are the input fields that you can update. Job ro. I'm going to use a solutions architectural ats as a job link.

I will provide a GitHub and website. Feel free to modify them, of course, with your own information. And let's kick off our Chrome. Here's how this process works. First, the job researcher analyzes the job link and extracts all the relevant information. Next, the personal profile scans my repo website to learn more about me. Then the resume strategist creates a new resume. And finally, the interview coach generates a set of interview questions. When this process runs, it creates two files. You can review and export them if needed. Now, let's look at the technical questions it generated.

I actually started at AWS as a solutions architect and went through a technical interview for this role. So, let's see how close these questions are. So one of them is walk me through your experience designing cloud solutions using AWS and it's actually a solid question that you will definitely get especially when they ask you to describe a complete cloud architecture. It's important to have a few examples and be ready to explain how you ensure scalability and security. I also like how it suggests some of my projects from my repo that I can use in references.

You can see it cover different categories from technical to strategic leadership and cross team collaboration. Well, now let's try something non AWS related. I found the solutions architect manager role position at OpenAI. I'm also going to switch foundation model to Amazon Nova which one of the newest models introduced by Amazon and it's significantly cheaper. There are three options micro light and prom and I'll go with prom. it's a multi model and it supports over 200 languages. Here is a quick comparison with code 3.5. It performs well on translation, instruction, following and agent orchestration.

It cost 80 cents per 1 million input tokens and about $3.20 per 1 million output tokens. We will use this numbers to estimate the total cost of running our crew. Now let's update the model name and the job link. So I can start this process. It should take about one or two minutes to run. This time it emphasizing my programming skills like Python and JavaScript and highlighting experience building business relationships with seale executives and leadership skills which are a lot more relevant for managerial role. So that kind of makes sense.

So let's check out the new resume. Okay, this is interesting. It says I have 14 AWS certifications. I don't even think AWS has 14 certification. And this is what known as an AI hallucination where the model just comes up with things and that's why it's always important to validate AI generated outputs. Looking at the summary of the resume, it focuses more on leadership and a IML experience, which makes sense for management position at an AI company. Now, let's check how much it cost to run this process. So I'm just going to multiply by the number of input and output tokens used.

So 80 cents per 1 million input, $320 cents per 1 million output and clothes onet cost $3 per 1 million input and $15 per 1 million of output tokens. So based on the task that we just run the total cost with Amazon no was around 3 cents well two 2.7 to be exact and with code it is about 11 cents. So both models performed well but since Nova is cheaper I would just keep utilizing the nova and refine my process. Now you can experiment try different models put your own inputs and see how the model responds.

So, it's really your turn now to experiment and feel free to check out the repo. I've linked this in the description. You will also find a tutorial there on investment and stock analysis which follows similar approach. You can use it as a foundation to start building your own AI agents. Please share with me what ideas you have and what you're going to build. I'm really curious. And if you find this helpful, feel free to share with others who might benefit. and happy building.