From Confusion to Clarity: My Journey into AI Agents

Learning has no limit. I’m still learning every day. But I want to share my journey and my approach to learning AI agents. Maybe you’ll get some ideas if you’re just starting out.

When I decided to learn and build AI agents, the first thing I did was read some papers. I found a GitHub repo with a collection of papers and started reading them. I understood some basics of AI agents, but honestly, I didn’t understand that much in the beginning.

By the way, before I even started with agents, I already had some background with RAG (Retrieval Augmented Generation). I knew a bit about vector databases like FAISS and ChromaDB, chunking strategies, and evaluation metrics like BLEU for RAG.

Now, back to my agent journey. I discovered CrewAI and tried to learn the basics. I started with small projects like finding stock prices. I also ran some open-source projects. It was useful, but I still struggled to understand the underlying parts.

At that time, I read a lot of documentation. I ran existing projects and slowly got a sense of how agents actually work.

Then I found LangChain, especially LangGraph, and I loved it. I watched videos, read the docs, and tried to understand concepts like state, graph, the ReAct mechanism, and how to implement human-in-the-loop.

I even took a short Hugging Face course on AI agents. At the same time, I started watching some DeepLearning.AI videos while building small agents on my own.

At first, I was very confused. I tried to solve every problem using agentic flow, even when a simple workflow would have worked.

Some Anthropic blogs really helped me a lot to clear up my confusion. For example: “Building Effective Agents” — super useful article.

After that, I experimented with adding more complex features, like a voice wrapper for my agent.

When I got a deeper understanding of AI agents, I realized I also needed backend knowledge to expose endpoints for the UI. That’s when I learned some FastAPI and discovered Pydantic. I spent a lot of time with Pydantic because it felt really powerful. I’m still learning, but now with FastAPI and Pydantic, I have better control over how my agent operates.

Later I felt something was missing. I had done OOP with C++ in university, but my Python OOP skills weren’t strong.

So I decided to go back and learn Python for AI and OOP from scratch. I took the AI Python for Beginners course from DeepLearning.AI by Andrew Ng. It covered basics like Python, packages, and APIs — really helpful for building my foundation.

Then I discovered MCP (Model Context Protocol). It’s an open-source standard for connecting AI applications to external systems. I tried using Postgres, Slack, and GitHub MCP connectors in my agents. I still have more to learn here, but I can already see how powerful it is.

One thing I’ve realized is that prompting and context engineering are super important in this journey.