Large-model (LM) agents are proliferating across domains, yet current systems remain ad-hoc pipelines without operating system (OS) -level guarantees for scheduling, memory, real-time responsiveness, and end-to-end security. Today's agent architectures resemble the pre-OS era of computing—a chaos of duplicated solutions lacking fundamental abstractions for resource management, isolation, and coordination. Existing frameworks (e. g. , tool-calling, Model Context Protocol, Agent-to-Agent messaging) address isolated aspects but lack a unified, security-by-design, latency-aware foundation suitable for enterprise and safety-critical deployments. This paper introduces the conceptual Agent Operating System (Agent-OS) as a computational substrate for agentic workflows. We present this not as a system fully realizable today, but as an architectural North Star to guide the next decade of agent infrastructure research. We propose a unified requirements specification encompassing functional requirements (lifecycle, memory, tools, orchestration, observability, safety, governance) and non-functional properties (reliability, scalability, interoperability, compliance, real-time, security), extended with explicit latency classes—Hard Real-Time (HRT), Soft Real-Time (SRT), and Delay-Tolerant (DT). An abstract layered architecture is defined, comprising Kernel, Services, Agent Runtime, Orchestration, and User layers, with cross-cutting concerns for security, governance, and observability. Following a system engineering requirements-driven methodology, we (i) trace historical roots and survey emerging 2025 systems, (ii) synthesize requirements with latency taxonomies, (iii) formalize Agent Contracts for portability and enforcement, and (iv) map these to the layered architecture. The Agent-OS provides a blueprint for scalable, interoperable, and trustworthy agent deployment for next-generation Agentic-AI-powered smart cities, autonomous systems, and enterprise AI - even as full realization may require years of collective research effort.
Anis Koubâa (2025) studied this question.