How to Train a Large Language Model for Programming

We're pretty much so much agency. What we choose, what we even choose to learn, we can be like, oh, I don't like physics anymore, I'm going to move over to ballet because I'm better at that anyway. Presumably you couldn't turn a large language model into an agent, just by plugging its output into a normal computer and its input from the readout from the computer. Are there other domains where we can narrow in or generate, like let the model generate useful training data, or find or use its training data?