How to build AI Agents using Open source tools

In a generic sense, we witness a constant penetration of AI into the traditional systems. These systems can include anything from software development methodologies to SaaS products and services. A longstanding discussion exists about whether AI should be considered a standalone product or merely a feature. Currently, with state-of-the-art (SOTA) models being released regularly, AI appears to be evolving into a product. However, fundamentally, AI is expected to serve as a feature that enhances both traditional systems and human capabilities, ultimately improving performance significantly.

But for AI to take this leap, it must come out of the shackles of being merely a language model and a next token predictor. It should get access to tools, be able to use them, call them to get some fresh on the fly, up-to-date information, reassess its strategies for task execution, critique new plans, and then carry them out. This leap is achieved by AI agents.

In this workshop, we will use Langgraph to build out AI agents. Langgraph operates on the principle of constructing a Directed Acyclic Graph (DAG), where nodes represent tools or systems used for data retrieval—such as APIs, document loaders, or even calls to larger language models for tasks like information extraction. The edges represent the information flow from nodes. The DAG helps in defining the business logic of the application we are building. Langgraph supports the context-aware async execution with streaming output out of the box.

In this workshop, we will take a specific example of building an AI agent using Langgraph that can call an API in the financial domain, extract and load the data into a mongodb database and build aggregation pipelines on top of it to get summarized answers for the user's questions. In this specific example, the API calling, mongodb database and summarization are all subagents.