Agent Operating System: A Reference Architecture for Distributed Agentic Systems
There's a new paper on arXiv (2608.03214) that introduces something called the Agent Operating System (AOS). It's a flexible framework meant for systems with distributed agents, without being tied to any specific vendor. The authors argue that large language models have evolved from just making predictions to becoming key components in broader systems that can reason, use various tools, access data, delegate tasks, and act on behalf of users or organizations. Currently, the tools available—like agent frameworks, workflow engines, and communication protocols—do improve performance but lack a unified operating framework. AOS aims to fill this gap by handling intent, choosing capabilities, and managing delegation and uncertainty. It includes two internal components: a Control & Governance Plane and another yet-to-be-specified plane.
Key facts
- Paper proposes the Agent Operating System (AOS), a reference architecture for distributed agentic systems.
- AOS is vendor-neutral and implementation-independent.
- AOS addresses governance of intent, capability selection, authority preservation, uncertainty control, runtime coordination, and action reconstruction.
- AOS contains two internal planes: Control & Governance Plane and another plane.
- The paper is available on arXiv with ID 2608.03214.
- The announcement type is 'new'.
- Large language models are now components of distributed systems that reason, invoke tools, retrieve external state, delegate tasks, and act on behalf of users.
- Existing technologies like agent frameworks and observability tools improve execution but lack a stable operating architecture.
Entities
Institutions
- arXiv