This article is what’s dumb:

The durable version is agents whose native language is precise, machine-facing state, with the warm, concise, human version generated only at the boundary.

Cool bro, we’ll take the natural lagauage machine, somehow make it generate machine language instead, and then magically convert “machine facts” into “human interpretation” only at the “rendering boundary” which is always very clearly defined, as this article points out when it mentions how sometimes humans read LLM output and sometimes they pipe it directly to other LLMs.

There is no formula that generates only true statements.