I'm not super sure if this is true (yet?). I think that these newer LLMs are trained on results (the agent got some code to run with minimal prompting), and not

I'm not super sure if this is true (yet?). I think that these newer LLMs are trained on results (the agent got some code to run with minimal prompting), and not on text. (I think this is called RLVR.)