Learning With an Agent in the Loop

An agent can build a project beautifully before you learn a thing.

I have always learned by building, but only when the thing made sense to me. Apple documentation and Objective-C stayed abstract until I found an App Store idea I actually wanted to make. Then I was in.

LLMs sharpen the risk. I can set up an Elixir project, give an agent my context, and get an impressive result quickly. But if it makes all the choices, what did I actually learn?

My approach is deliberately slow: tiny chunks, explanations, and a suggested review order. I read the important code and ask a ridiculous number of questions—about alternatives, unfamiliar conventions, and how the same problem would look in a stack I know.

It is not optimized for speed. It is optimized for staying in the loop.

That is how I have learned TypeScript, Elixir/Phoenix, and Rust lately: small real products with no monetary goal, built for fun. Writing code this way feels a lot like video games to me: challenge, progress, no pressure to ship.

The agent writes with me. I still do the learning.