Building an AI agent with langGraph (step by step tutorial)

Key Takeaways

Hi, I'm Annie with the MLOps community and I am going to talk about how to pick a topic and build your first agent. So, if you've been hearing about agents and you don't know how to get started, but you'd like to dip your toes in the water, um this is a great start start tutorial for that. One of the first things that you need to do to develop an agent is find a problem that requires an agent framework. Even though you can incorporate rag and chat bots into your agent, one of the most important features is that you're giving the AI specific roles and those roles are making decisions at different points in your framework. There are many different frameworks for building agents but for the purpose of this tutorial I chose langraph. So langraph uses a graphbased approach to orchestrate workflows. Langraph also allows for fine grained control of your agent and it also gets you to a place where you could have an agent that is ready to be deployed. So I liked both of those features. That's why I chose to build my agent in Lingraph. First, I took the Langraph Academy course called Intro to Langraph. It's actually free at the time of recording this video, which is super useful and resourceful. Now, you can watch the entire training, but I would definitely focus on modules 1 through 4 just to get yourself started. These modules build upon each other and show you different features that Lang graph offers in order to be able to build an agent such as what states are in an agent, how to incorporate memory into your agent, and then also how to incorporate something like human in the loop. The module that I would definitely focus on once you make it through the videos in one through three is module four and especially this research assistant because Lang graph academy gives you this really great example of a an agent that you can use as a baseline or foundation to create your own agent and really understand what the agents are doing and how the code works. works. So then I reviewed the code from the langraph tutorial that built the research assistant AI agent. This was really cool because they set up several different roles like assistants and experts to basically interview each other on topics and then write an academic paper. So, this was a great um example of an agent and the code is really clearly laid out once you've sort of gotten familiar with the structure of langraph. I ended up deciding to take pieces of this code, especially the parts where the nodes and edges are built, and then build my agent off of that example code. So then I needed to think of, okay, what do I want my agent to do? I came up with an idea that would require the agent to perform several tasks, not just something that, you know, a chatbot could answer on the fly. And then I also wanted to use a data source that wasn't publicly available so that the the agent would have to get to know some data that it wasn't trained on. So, I chose my credit card data, which is thankfully not publicly available. I also wanted the agent to do something silly. So, I decided that I would create an agent that is a passive aggressive budget coach that is coaching me on all the credit card purchases that I make in my statement. So with a little bit of help from cursor, I took the research assistant agent code and then I sort of morphed it into a different agent that was focused on coaching me on my budget. As you can see, this is set up very similarly to the code from the research assistant. So first I I create the classes for the nodes that are going to be in my agent such as grabbing the transaction data um analyzing the transaction data and then interacting with the budget coach through using LLMs. Then I defined what the budget coach would do. So, pulling down transactions, analyzing them, and then looking at them against my monthly budget, and then coming up with some snarky comments about where I'm at with my budget thus far. I also determined that the agent would alert me when I reached a certain percentage of my budget. So lastly, I set up the nodes and edges of the agent so that the workflow of the agent would flow from analyzing the transactions, generating the summary for the transactions and then sending an alert. Langraph has this really nice feature called Langraph Studio where you can actually view the graph structure of your agent. So this sort of outlines the start of the agent. It analyzes your transactions, generates a summary, and then it ends. And it will send an alert based on certain conditions. In this case, if I've spent up to 80% of my monthly budget, you can check out my code for this in a GitHub repo that is posted in the description of this video. One other quick thing, credit card companies won't provide individuals with APIs to access their credit card data. So, I subverted this by using Plaid, which you can basically use Plaid to log into your credit card account and pull data from your statements that way. Plaid has a really convenient Python library that allows you to do this. And the thing that you need for it is first of all to get a plaid account and then to log in and uh basically grab their API key which will allow you to be able to log into your bank account. For the purpose of this tutorial, I'm just using the sandbox environment in Plaid, which contains fake credit card data. Uh, which that's where this username and password is for. Um, you can grab API keys under this keys section, which I won't do right here because it will reveal my keys. But there's a key for both the sandbox environment and the production environment. and the production environment will allow you to get your actual credit card data from your accounts and you can log into that. I didn't want to put my credit card data on this video. So, uh, for now we're using some fake data. So then I set up a a specific script that uses the Plaid Python library to pull down transactions for the past 30 days and then it categorizes those transactions based on the descriptions in the transactions. And this feeds back to the agent itself. Lastly, I really wanted to be able to see my agent in an application so that I could sort of maybe use a little bit of human in the loop with it. So, I created a Streamlit app that made this budget coach come alive and that's what's running right here. So, if you can see now that I'm running this locally, um, okay, welcome to your passive your passive aggressive budget coach. I'll analyze your spending habits and provide some constructive feedback. So, great. And so, I have this set to 5,000 because I want for the purpose of this example to show you what it does when you're at 80% of your budget. Um, so this is using fake data. You can click here and it will show a spending summary. Um, this takes a little bit to run, but it also sends out this alert to my email. And this says, uhoh, spending alert. Your budget was 5,000, but you're almost reaching that 5,000, and this is how much you have remaining. And then it gives its uh sarcastic kind of snarky comments about things that I've spent money presumably on. Apparently I spent $300 on a pair of sneakers. So yeah, I think this is kind of the overall goal of the budget coach. I think my next steps for this agent would be to use more of the human in the loop features that Langraph offers to just play around with changing states and also to deploy it on more than just a local machine. Now that you've seen me walk through setting up my own agent, you can try this for yourself. go through the Lang graph tutorial or just skim through it if you don't have a ton of time and then use some of the langraph code available in this tutorial to think of your own idea and just spin up a prototype of your your first agent. And stay tuned in to the MLOps community for more rad content on agents.