Interesting idea! I wonder if, after some iteration, a variation on this could help curb some of the LLM spaghetti mess