arXiv paper locks AI agents to rules they cannot skip at run time

A research paper proposes a framework where a contract binds an AI agent at the moment it acts. Before any action runs, the agent must produce a mathematical proof that the action satisfies the contract. If the agent cannot produce that proof, the system blocks the action. This matters for designers because it moves trust from assumption to evidence. Today, a human who grants an agent access to sensitive tools must assume the agent respects stated limits. This framework refuses unproven actions before any harm occurs. The human sets the terms once. The framework enforces them without requiring the human to watch. The approach also covers chains of agents, where one agent hands tasks to another. A contract travels with the work, and each agent must satisfy it before acting.