Agent Operating System Proposed for Distributed Agentic AI Systems

The rapid evolution of large language models into complex, distributed agentic systems has created a need for a more robust and standardized operating architecture. Current frameworks and tools improve execution but lack a unified system for managing intent, authority, uncertainty, and coordination across these agents. The Agent Operating System (AOS) is introduced as a vendor-neutral reference architecture to fill this gap.AOS is structured into two main planes: a Control & Governance Plane, which handles aspects like intent, policy, trust, and human oversight, and a Runtime & Coordination Plane, responsible for agent lifecycle, workflow, model routing, and scheduling. This architecture aims to provide a stable foundation for building governable, reliable, observable, and interoperable agentic systems, integrating with existing infrastructure through explicit interfaces. The paper outlines AOS concepts, invariants, and responsibilities, while also identifying open research questions.