Behavior-Constrained Agent Systems: A Control-Theoretic View of Safe Authorization

This paper presents a systems and control-theoretic framing for behavior-constrained authorization in AI agent systems. Building on behavior-bound signature ideas introduced by Y.Y.N. Li, it argues that high-risk agent actions should be governed not merely by stronger signatures, but by a negative-feedback control architecture in which actions are structured, authorization is policy-bound, validators act as mandatory execution gates, and structured rejection reasons serve as error signals for safe retry. The paper develops a reference architecture, an action-modeling discipline, and application analyses covering payments, API invocation, autonomous software development, database mutation, file deletion, and enterprise approvals. It argues that blockchain is not required for the first useful deployments, and that behavior-constrained authorization is most valuable today as a hard control layer for high-risk machine actions.