Treating LLM as the CPU: Toward a General-Purpose Agentic Computing Specification and Operating System Architecture

Abstract

Current application paradigms for Large Language Models (LLMs) remain confined to simplistic "wrapper" encapsulations, lacking system-level memory management, standardized communication protocols, and auxiliary intelligence frameworks. This paper analogizes the LLM to the "CPU" of a novel computing architecture and, from this perspective, develops a vision for an "Agent Operating System" (Agent-OS) and a general-purpose agentic computing specification. We propose a comprehensive analogical framework: virtual context management mechanisms (MemGPT/Letta), memory layer architectures (Mem0), and five-layer self-learning memory systems (Hermes Agent) constitute the "RAM/Disk" subsystem; the Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication protocols form the "Bus"; tool APIs serve as "Peripherals"; and application frameworks such as OpenClaw's gateway architecture and Hermes Agent's self-evolutionary learning loop function as "Applications." At the deployment level, we analyze the cloud-edge-device three-tier collaborative computing paradigm, elucidate the technological pathways and industry practices for "on-device personal agents," and envision the interconnection prospects of the "Internet of Agents." Furthermore, we explore the self-evolution capabilities of AI agents—from experiential memory accumulation to self-optimization, and ultimately to meta-cognitive cross-domain self-improvement—revealing a paradigm shift from "fixed programs" to "living code." Finally, we outline the standardization framework required for a general-purpose intelligent application specification, encompassing a unified memory file system, a standardized bus protocol stack, and an AI operating system kernel with SDK specifications, thereby providing a conceptual framework and architectural blueprint for systematic research in this domain.

Keywords