Some replies along the lines of the key idea is breaking down the prompt and recursively running sub LLMs on it and stitching them back together.
I understand that’s the central premise of the paper. But the lasting nugget is giving the LLM a REPL (which may not even be novel in this paper).
Prescriptive recipes for how to do something will almost certainly not make sense in the near future, maybe even in the next generation of LLMs. They will be washed away in the tide of increasing LLM capabilities.