Title: Model-Native Computing Architecture: Envisioning Future System Architecture Through the Lens of Computer Architecture

Abstract

Large language models are undergoing a significant shift, evolving from mere technological components into foundational system technologies. As developers increasingly leverage tools such as Codex, Claude Code, and AutoGPT to automate coding, project management, and complex multi-step workflows, they encounter recurring engineering challenges. Issues regarding cache reuse, context management, agent scheduling, and permission control now closely mirror classical computer systems problems. This paper explores this analogy in a visionary survey, mapping traditional computer architecture concepts onto the emerging model-native stack. We examine existing research on LLM-as-OS paradigms, memory management, agent frameworks, tool protocols, multi-agent coordination, cognitive architectures, and safety governance.

Our analysis suggests that while these diverse research strands address different layers of the same system, they currently lack a unified theoretical model. To bridge this gap, we propose the Intelligent Computing Architecture Model (ICAM), a comprehensive six-layer framework designed for model-native computing. ICAM establishes explicit interface contracts and design axioms to provide structural clarity. Furthermore, the framework resolves the debate regarding whether an LLM functions more like a CPU or an operating system by introducing a dual-plane perspective. This view distinguishes between a probabilistic execution plane, which focuses on computational possibilities, and a deterministic control plane, which governs computational intent.

We also define three fundamental design laws: the Semantic Locality Law, which addresses KV-cache reuse and inference acceleration; the Context Budget Law, which manages effective working sets within finite windows and under attention decay; and the Agent Speedup Law, which highlights diminishing returns in multi-agent collaboration. These laws are validated against published system-level data and correlated with recent findings on agentic software practices. Finally, we identify the limitations of the computer architecture analogy and outline a research roadmap for the field of model-native computing. This work serves as a conceptual and survey contribution and does not present new experimental results.

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