Title: Agent Operating Systems (AOS): Embedding Agentic Control Planes Within and Beyond Conventional Operating Systems

Abstract:

Conventional operating systems were architected around deterministic software, explicit control flows, and workflows initiated by users. Their fundamental abstractions—including processes, threads, system calls, files, and permissions—rely on the assumption that behaviors are bounded and interactions follow predictable patterns. In contrast, agentic AI systems operate under a distinct execution model characterized by long-lived, goal-oriented entities. These agents reason probabilistically, dynamically invoke tools, and adapt their actions in response to feedback. Although current implementations of agents exist as user-space applications, their operational traits place significant strain on traditional OS boundaries, particularly regarding scheduling, memory and state management, security, observability, and governance.

This study proposes the Agent Operating System (AOS), a systems architecture designed to embed an agentic control plane into existing operating systems or, in certain configurations, gradually assume specific OS responsibilities. We offer a rigorous definition of AOS, outline explicit assumptions and limitations, and systematically decompose AOS functions into five key areas: schedulers, context and memory management, tool and capability registries, policy and trust enforcement, and observability and audit.

The paper examines the inadequacies of classical OS abstractions when applied to agent workloads and explores various integration strategies, ranging from user-space runtimes to distributed control planes. Furthermore, it maps AOS concepts onto the primitives found in Linux and Windows environments. We also address security and safety considerations, detailing agent-specific threat models and establishing evaluation criteria that prioritize deterministic enforcement, auditability, and clarity for operators. The primary aim is not to entirely supplant operating systems, but to create a robust systems foundation for agentic computation that ensures controllability, accountability, and security at scale.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC