super interesting how it makes "decisions", but nice that they let you tie user feedback directly into LLM refinement, otherwise would be hard to make that info useful