隨著嵌入式物聯網系統逐漸應用於工業控制、機器人與即時分散式環境，DDS (Data Distribution Service) 雖能提供可配置服務品質與高可靠通訊能力，但其開發需處理多項服務品質策略、通訊生命週期、資料序列化與平台相依設定，導致傳統人工開發流程複雜且容易產生配置錯誤。為提升 DDS 中介軟體程式生成之可控性與可驗證性 ，本研究提出 MIAT-DDS 多 AI (Artificial Intelligence) 代理框架，結合LLM (Large Language Model) 與 MIAT 系統設計方法論，將自然語言需求依序轉換為需求規格、IDEF0 系統架構模型、GRAFCET 離散事件模型與可部署之 XRCE-DDS (DDS for eXtremely Resource Constrained Environments) C 語言程式碼，並透過需求定義代理人、系統架構設計代理人、離散建模代理人與高階合成代理人建立具架構約束與行為約束的鏈式協作流程。實驗結果顯示，階層化中介表示可有效降低 LLM 生成過程中的幻覺與結構不一致問題，並為嵌入式 DDS 中介軟體之自動化生成、部署驗證與效能調校提供具工程可追溯性的系統化方法。;As embedded Internet of Things systems are increasingly applied to industrial control， robotics， and real time distributed environments， Data Distribution Service (DDS) can provide configurable Quality of Service and highly reliable communication capabilities. However， its development requires the handling of multiple QoS policies， communication lifecycles， data serialization， and platform dependent configurations， making conventional manual development processes complex and prone to configuration errors. To improve the controllability and verifiability of DDS middleware code generation， this study proposes MIAT-DDS， a multi Artificial Intelligence (AI) agent framework that integrates Large Language Models (LLMs) with the MIAT system design methodology. The proposed framework sequentially transforms natural language requirements into requirement specifications， IDEF0 system architecture models， GRAFCET discrete event models， and deployable DDS for eXtremely Resource Constrained Environments (XRCE-DDS) C language code. It further establishes a chained collaborative workflow with architectural and behavioral constraints through four agents， namely the requirement definition agent， system architecture design agent， discrete modeling agent， and high level synthesis agent. The experimental results show that hierarchical intermediate representations can effectively reduce hallucination and structural inconsistency during LLM based generation. They also demonstrate that the proposed framework provides a systematic and engineering traceable approach for automated generation， deployment verification， and performance tuning of embedded DDS middleware.
