Mesa-optimizers are typically described as internal agents that emerge unpredictably inside trained models. This paper argues that they are neither mysterious nor unique to machine learning. Instead, mesa-optimizers are the outcome of classical principal-agent problem geometry operating inside recursive optimization architectures. When a learning system reinforces internal routines on the basis of imperfect proxies, substructures that capture reward correlations become recursively entrenched, amplifying their influence and drifting from designer intent. This incentive geometry – local proxies, partial observability, asymmetric reinforcement, and recursive feedback – is structurally analogous to that seen in economics, biology, and organizational behavior. Reframing mesa-optimizers as principal-agent distortions clarifies their origin and suggests mitigation strategies analogous to those used in other complex adaptive systems.