This paper addresses the growing trend of large-scale language model (LLM)-based agents that autonomously interact, collaborate, and delegate tasks without human intervention. Existing agent system designs are theoretical, lack concrete implementations and evaluations, and lack user controllability. Therefore, this paper proposes SAGA, a scalable agent system governance security architecture that provides user oversight. SAGA supports agent registration, user-defined access control policy management, and policy enforcement for inter-agent communication through a central entity, the Provider. Furthermore, it introduces a cryptographic mechanism for fine-grained control of agent interactions, providing formal security guarantees. Evaluations on multiple agent tasks using on-device and cloud LLMs demonstrate minimal performance overhead and no impact on the underlying task utility.
