The article presents a formal approach to trust calibration in automated agents, focusing on when these agents can act independently versus when they need human approval. This method leverages preference learning and Bayesian optimization to enhance decision-making efficiency in agentic tool use.