Do you have thoughts on the code-style agents recommended by huggingface? The pitch for them is compelling, since structuring complex tasks in code is something very natural for LLMs. But then, I don’t see as much about this approach outside of HF.