In my experience LLMs tend to work far better when they have a tree of structured information to search through and change.

Compared to poking at a mungy ball of accidental state where every change introduces complexity for the future which every downstream agent needs to correctly find.