Restricting an agent's decisions to obey real-world physical laws and safety limits, preventing it from proposing impossible or dangerous actions.