A Four-Layer System Architecture for Agentic AI: An Integrated Framework from Cross-Literature Synthesis
宋嘉祺1
作者信息
1. 河南科技大学
折叠
摘要
[目的] 针对大型语言模型（LLMs）从被动文本生成器演化为主动推理驱动的自主智能体（Agentic AI systems）后，学术界与工业界缺乏统一架构描述框架的问题，提出一个整合三篇核心研究的概念框架。
[方法] 基于三篇核心文献——Arunkumar et al. (2026)的《Agentic Artificial Intelligence (AI): Architectures, taxonomies, and evaluation of large language model agents》、Jiang et al. (2026)的《SoK: Agentic Skills — Beyond tool use in LLM agents》、Chen et al. (2025)的《Towards reasoning era: A survey of long chain-of-thought for reasoning large language models》——采用逆向工程与概念综合方法，通过定义公理体系、层间接口契约和映射准则，提取并形式化定义四层系统架构模板。
[结果] 提出并形式化定义了四层系统架构：基础设施层（Infrastructure Layer）、认知推理层（Cognitive Reasoning Layer）、技能与工具整合层（Skill & Tool Integration Layer）、对齐与治理系统（Governance & Alignment Layer）。该架构构成了"算力-思维-行动-约束"的完整闭环。对每一层给出了严格的定义边界、核心构成要素、来源文献映射以及层间接口契约，并针对每一层可能的关键失败模式提出了防御性设计策略。
[局限] 框架目前为理论综合与形式化推导，尚未在大规模真实Agent系统中进行实证验证。
[结论] 本框架为Agentic AI系统的设计、分析与比较提供了统一的概念基准。
Abstract
[Objective] As large language models (LLMs) evolve from passive text generators to actively reasoning autonomous agents (Agentic AI systems), both academia and industry lack a unified architectural description framework. This paper proposes a conceptual framework integrating three core research works.
[Methods] Based on three core references—Arunkumar et al. (2026), Jiang et al. (2026), and Chen et al. (2025)—this study employs reverse engineering and conceptual synthesis to extract and formalize a four-layer system architecture template.
[Results] A four-layer system architecture is proposed: Infrastructure Layer, Cognitive Reasoning Layer, Skill & Tool Integration Layer, and Governance & Alignment Layer, forming a complete "Compute–Cognition–Action–Constraint" closed loop. For each layer, strict boundary definitions, core components, source literature mappings, and inter-layer interface contracts are provided, along with defensive design strategies against critical failure modes.
[Limitations] The framework is currently a theoretical synthesis without empirical validation in large-scale real-world agent systems.
[Conclusions] This framework provides a unified conceptual benchmark for the design, analysis, and comparison of Agentic AI systems.
Keywords: Agentic AI, system architecture, long chain-of-thought, agentic skills, AI alignment, Model Context Protocol
国家预印本平台