Towards Agents That Know When They Don't Know: Uncertainty as a Control Signal for Structured Reasoning