DEV Community

Cover image for I Stopped Chasing AI Frameworks and Started Building with Plain APIs
Caelora
Caelora

Posted on

I Stopped Chasing AI Frameworks and Started Building with Plain APIs

When I first got into AI development, I thought I needed to learn every new framework that appeared on GitHub.

LangChain.
CrewAI.
AutoGen.
LangGraph.
LlamaIndex.

Every week, there seemed to be another library promising to make AI agents easier.

I kept bookmarking repositories instead of actually building anything.

Eventually, I decided to ignore all of them for a weekend.

Instead, I opened the OpenAI API documentation and started from scratch.

The surprising part

The first chatbot I built was around 60 lines of code.

Adding conversation history wasn't difficult.

Calling a weather API wasn't difficult.

Generating structured JSON wasn't difficult either.

Suddenly I realized something:

Most AI frameworks aren't introducing new capabilities—they're packaging patterns that you can learn yourself.

That realization completely changed how I approached AI projects.

Frameworks aren't the enemy

Don't get me wrong.

Frameworks are useful.

They solve repetitive problems.

They standardize workflows.

They save time once your project grows.

The problem is when beginners mistake the framework for the technology.

It's like learning React before understanding JavaScript.

Or learning Kubernetes before understanding Docker.

You'll eventually hit a wall because you never learned the fundamentals.

What every beginner should build first

Before touching any AI framework, I'd recommend building these yourself:

  • A basic chatbot
  • Conversation memory
  • Function/tool calling
  • Streaming responses
  • JSON structured outputs
  • A simple RAG app
  • A tiny agent loop

None of these require thousands of lines of code.

Most can be built in an afternoon.

Why this matters

When something breaks inside a framework, you'll know why.

When a model behaves strangely, you'll understand where the prompt lives.

When tokens get expensive, you'll know exactly what's consuming them.

And when a new framework becomes popular next month...

...you'll learn it in hours instead of weeks.

Final thought

The AI ecosystem moves incredibly fast.

Frameworks will come and go.

APIs will change.

Model names will change.

But understanding the fundamentals will always be valuable.

Sometimes the fastest way to learn isn't by installing another dependency.

It's by writing one more file from scratch.

Top comments (0)