一句话：提出 LLM Agent 系统架构应围绕控制、知识、责任、约束与演化五个结构维度，而非仅关注组件与连接件。
论文指出，LLM Agent 的许多关键系统行为并非来自硬编码逻辑，而是由模型推理、策略选择与运行时状态涌现；传统架构图常遗漏提示、记忆与策略，导致对系统行为的描述不完整。作者主张，LLM Agent 架构不能只关注组件与连接件，还需架构“行为空间”，并从控制结构（谁决定下一步）、知识结构（记忆如何组织与治理）、责任结构（谁对什么负责）、约束结构（哪些行为被允许）和演化结构（系统如何随时间变化）五个维度进行推理。文章不提供具体解决方案，而是给出一个用于讨论、辩论与开发这些架构基础的坐标系与词汇表。
In LLM agent systems, many critical system behaviors no longer originate from hard-coded logic but emerge from model inference, strategy selection, and runtime state. This raises a fundamental question: when behavior is not in the code, what are we actually architecting? Traditional architecture diagrams often omit prompts, yet prompts determine system behavior. They often omit memory, yet memory determines long-term system drift. They often omit policies, yet policies determine whether systems remain safe. We may be drawing incomplete diagrams. This paper argues that LLM agent architecture requires reasoning about five structural dimensions: control structures (who decides what happens next), knowledge structures (how memory is organized and governed), responsibility structures (who is accountable for what), constraint structures (what behaviors are permitted), and evolution structures (how systems change over time). While components and connectors remain foundational, they may be insufficient to fully characterize these systems; we must also architect the behavior space. We offer not a solution but a coordinate system—a vocabulary to discuss, debate, and develop these architectural foundations.