ABSTRACT The rapid advancement of large language model (LLM) technology is profoundly transforming the practice of social science research. Scholarly discussions on Artificial Intelligence (AI)'s role in social science research can be organised into three levels: AI as a research tool, AI as a methodological infrastructure and AI as a quasi‐cognitive actor. Existing research studies predominantly focus on individual levels, with limited attention to how the tool and infrastructure levels interact and mutually support each other. Drawing on 2 years of practical exploration, this paper proposes a systematic framework spanning both the tool and infrastructure levels and demonstrates the bidirectional interaction mechanism between these levels through the development and operation of over 40 application systems. The framework comprises three tiers: the ontological tier (five core principles), the methodological tier (a technical architecture of ‘one core, three repositories, four domains’) and the practical tier (over 40 application systems). Top–down guidance and bottom–up feedback mechanisms among the three tiers form a continuously evolving closed loop. As the nexus connecting infrastructure and tools, this paper distils POMASA (pattern‐oriented multi‐agent system architecture), a methodological framework containing 20 design patterns that can guide the rapid construction of declarative multi‐agent research systems. The case studies provide detailed accounts of representative applications, including the digital sovereignty index (DSI) assessment system and the report production system (RPS), demonstrating the framework's operation across research tasks of varying scales and domains. This paper's contributions lie in demonstrating the bidirectional interaction mechanism between tools and infrastructure in AI‐assisted social science research, distilling a reusable methodological framework, and validating the framework's feasibility through extensive case studies.
Jie Xiong (2026) studied this question.