LLM agents are writing English (…) documents, ostensibly for human review and consumption, but 90%+ of what I have my LLM agents write is exclusively consumed by other agents, … and accessed by other agents - it seems like the English translation is pretty useless for that layer …

But remember, the LLM is trained on human language. It doesn’t understand what it is talking about, it just simulates humans talking about it. I don’t think there is a more fundamental token stream there to be uncovered.