I like the idea of llms collaborating like this a lot; planning, critiquing, verifying, coding etc. I think that’s a very general and powerful approach. How did you end up with that structure and what did you try first? What are the downsides? How do the component agents communicate, just json?