A technique I’ve started using for learning complex things: “learn by pull request”, or “learn by commit”.
The idea is simple. I get an agent to do something for me - test out a technique, try an implementation of a library, build a small thing - and then I learn by watching it work. Doing the deploys, seeing the errors, reading the diffs. It’s how I used to learn anyway, just done manually, where elapsed time was far longer.
Doing it all by hand was deeper in one sense - you can’t avoid learning something when you’ve typed every line yourself. But realistically you can’t get as much done that way; time constraints are real. What I’ve found is that agent-led work tends to be more comprehensive too, because the agent considers things I wouldn’t have thought of, which deepens the learnings again. Net-net: a better result overall, in a fraction of the time.
I’ve got a pile of these repos lying around - I’ve done many by now - and usually keep them that way. They’re notebooks more than projects.
Here’s a recent one. I call it “kodr” - as in coder. It’s a learning repo focused on a simple AI code harness. Also zero node dependencies. Why?!
It’s one thing to read about agentic harnesses, it’s a whole bigger step to try writing one. You have to take a lot of things into account. I also focused on using local models for this, via LM Studio. Why again?! Using a local model forces you to think hard about context, tools support, permissions, splitting work - all the stuff that gets papered over when you’re hitting a hosted frontier model.
Link to the repo below. Keep in mind it’s a learning tool - not a coding harness competitor!
Links: