A few months ago I asked why semantic representation rather than text wasn't used, since natural language seems quite a lossy representation for semantic concepts:
https://news.ycombinator.com/item?id=47195212
I wouldn't have thought to use it for LLM-to-LLM communication, though