Agent Operating System: Standardizing Abstractions for AI Agents.

The evolution of platform software typically follows a pattern: initial experimentation, followed by the definition of stable abstractions, and finally consolidation around these to create portable platforms. This pattern was observed with POSIX for classical operating systems and Kubernetes for cloud environments. The paper posits that agentic AI systems, which are autonomous, LLM-driven entities capable of planning, tool use, memory, and collaboration, are currently in the early, experimental phase of this cycle. With numerous frameworks and protocols emerging, there is a lack of community consensus on core abstractions and their guarantees, hindering portability and reliable composition of agentic applications. The authors advocate for a path forward that mirrors prior waves: extending classical and cloud OS primitives to accommodate the stochastic, natural-language-mediated execution inherent in AI agents. By precisely specifying the semantics of these new agentic abstractions, the field can move beyond prototypes towards a consolidated platform, enabling more robust and scalable agentic AI development.