隨著科技製造業資訊系統高度碎片化，工廠內部常形成嚴重的「資料孤島」。傳統 RPA 多依賴預先設定規則與固定流程，較難處理需情境判斷之任務，無法解析交接班日誌等非結構化數據，導致跨系統異常診斷高度依賴工程師的個人經驗，平均修復時間 (MTTR) 居高不下。為迎向「代我執行(DIFM)」的產業趨勢，本研究旨在建構一套專為科技業設計的 Agentic AI（代理式人工智慧）系統架構，以重塑跨系統資訊整合與除錯流程。
本研究結合「系統規格設計」與「專家深度訪談」雙軌方法。首先透過現況(AS-IS)診斷拆解異常診斷瓶頸，設計 TO-BE 智慧化流程；接著邀請具備 IT 維運、AI 架構與資安風控經驗的業界專家進行半結構化訪談，從技術可行性、業務契合度與風險管控等構面進行實證驗證與架構優化。
在架構設計上，本研究提出五大層級的「主控代理-專業代理(Supervisor-Worker)」多代理系統(MAS)。主控代理負責意圖識別與任務拆解，並分派給負責處理結構化資料的「製程數據代理」、分析時序資料的「設備狀態代理」，以及利用 Agentic RAG 技術萃取非結構化文本脈絡的「脈絡檢索代理」。此架構透過去中心化推論與 API 閘道器，有效解決單一大型語言模型(LLM)的幻覺與運算瓶頸。專家訪談亦證實此設計能對齊跨系統數據，預估可縮減異常診斷時間。
本研究透過系統化的架構規格設計，初步建立Agentic AI應用於複雜工業場域的概念框架，並藉由專家訪談取得初步效度支持。確立科技業導入 AI 時必備的「人類在迴圈內(HITL)」確定性安全護欄。本研究可作為科技製造業推動智慧工廠、強化資訊整合與提升營運效能之參考依據。

As information systems in the technology manufacturing industry become highly fragmented, severe "data silos" frequently emerge within factories. Traditional Robotic Process Automation (RPA) can only handle deterministic tasks and is incapable of parsing unstructured data such as shift handover logs. Consequently, cross-system anomaly troubleshooting relies heavily on the personal experience of engineers, leading to a persistently high Mean Time to Repair (MTTR). To embrace the industry trend of the "Do-It-For-Me" (DIFM) paradigm, this study aims to construct an Agentic AI system architecture tailored for the technology sector to reshape cross-system information integration and troubleshooting workflows.
This research adopts a dual-track methodology combining system specification design and in-depth expert interviews. Initially, an AS-IS diagnosis is conducted to deconstruct troubleshooting bottlenecks, followed by the design of a TO-BE intelligent workflow. Subsequently, industry experts with extensive experience in IT operations, AI architecture, and cybersecurity risk management were invited for semi-structured interviews. These interviews serve to empirically validate and optimize the architecture from the perspectives of technical feasibility, business alignment, and risk management.
In terms of architectural design, this study proposes a five-layer "Supervisor-Worker" Multi-Agent System (MAS). The orchestrator agent is responsible for intent recognition and task decomposition, dispatching tasks to specialized agents: an "MES Agent" handling structured data, an "EQP Agent" analyzing time-series data, and a "Context Agent" utilizing Agentic RAG technology to extract contextual information from unstructured text. Through decentralized reasoning and an API gateway, this architecture effectively mitigates the hallucination and computational bottlenecks associated with a single Large Language Model (LLM). Expert interviews further confirmed that this design can successfully align cross-system data and is projected to significantly reduce troubleshooting time.
Through systematic architectural specification design, this study establishes a preliminary conceptual framework for applying Agentic AI in complex industrial scenarios, obtaining initial validity support via expert interviews. It solidifies the indispensable "Human-in-the-Loop" (HITL) deterministic safety guardrails required for AI deployment in the technology industry. By proposing differentiated deployment strategies for enterprises of varying scales, this research provides an empirically grounded architectural blueprint for the technology sector to unlock operational leverage and advance toward smart factories.

1. Association for the Advancement of Artificial Intelligence (AAAI). (2025). AAAI 2025 Presidential Panel on the Future of AI Research.
2. Blount, A., Gulli, A., Saboo, S., Zimmermann, M., & Vuskovic, V. (2025). Introduction to Agents.
3. Coshow, T., Gao, A., Pingree, L., Verma, A., Scheibenreif, D., Khandabattu, H., & Olliffe, G. (2024). Top strategic technology trends for 2025: Agentic AI. Gartner.
4. Gulli, A., Nigam, L., Wiesinger, J., Vuskovic, V., Sigler, I., Nardini, I., Stroppa, N., Kartakis, S., Saribekyan, N., Nawalgaria, A., & Bount, A. (2025). Agents companion. Google.
5. Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., ... & Wang, H. (2023). Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997.
6. Ghose, R., Shah, R. S., Bantanidis, S., Chahal, S., & Master, K. (2025). Agentic AI: Finance & the 'Do It For Me' Economy. Citi Global Perspectives & Solutions (Citi GPS).
7. Hong, S., Zheng, X., Chen, J., Cheng, Y., Zhang, C., Wang, Z., ... & Wu, C. (2023). MetaGPT: Meta programming for a multi-agent collaborative framework. arXiv preprint arXiv:2308.00352.
8. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., ... & Yih, W. T. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459-9474.
9. Liu, Z., Yao, Y., Li, J., Ye, J., Wang, X., Hai, C., ... & Sun, M. (2023). AgentBench: Evaluating LLMs as agents. arXiv preprint arXiv:2308.03688.
10. Madkour, N., Newman, J., Raman, D., Jackson, K., Murphy, E. R., & Yuan, C. (2026). Agentic AI Risk-Management Standards Profile (Version 1.0). Center for Long-Term Cybersecurity, UC Berkeley.
11. Packer, C., Fang, V., Patil, S. G., Lin, K., Campanella, S., Gonzalez, J. E., & Gonzalez, R. (2023). MemGPT: Towards LLMs as operating systems. arXiv preprint arXiv:2310.08560.
12. Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology.
13. Patel. (2019). Bridging Data Silos Using Big Data Integration
14. Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., ... & Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.
15. Tapscott, D., & Caston, A. (1993). Paradigm Shift: The New Promise of Information Technology.
16. Wang, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., ... & Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.
17. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., ... & Wen, J. R. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6), 186345.
18. Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., ... & Wang, H. (2023). AutoGen: Enabling next-gen LLM applications via multi-agent conversation. arXiv preprint arXiv:2308.08155.
19. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing reasoning and acting in language models. International Conference on Learning Representations (ICLR).
20. Zheng, L., Chiang, W. L., Ying, Y., Shen, S., Anyscale, S., Lin, Z., ... & Xing, E. P. (2024). Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. Advances in Neural Information Processing Systems.
21. 麥肯錫全球研究院 (McKinsey Global Institute). (2023). 生成式 AI 的經濟潛力：下一個生產力新紀元 (The economic potential of generative AI: The next productivity frontier). McKinsey & Company.