With the development of artificial intelligence (AI), especially large language models (LLM), LLM- driven dialogue systems have begun to exhibit powerful intelligence and autonomy. They not only make humans feel as though there are social relationships when communicating with non-human entities, but also blur the boundary between the real world and virtual worlds, especially those related to video games. This thesis presents an innovative exploration into the development and understanding of AI agents as hybrid agents, specifically focusing on LLM-driven conversational agents within interactive story- telling. It aims to bridge the gap between fictional narratives and social interactions in digital spaces, leveraging LLMs to enhance the believability and engagement of these agents. To validate and explore this framework, a series of practice-led studies were conducted. These artifacts—including the Wander chatbot based on real-world map, AI-native game 1001 Nights, the community-integrated chatbot Catherine & David, the creativity support tool ORIBA, the conceptual art installation AI Nu ̈shu, and the final integrated system Hyborg Agency —were evaluated through a mixed-methods approach combining quantitative data with in-depth qualitative analysis. A key finding of this thesis is a more nuanced conceptualization of believability. The studies re- veal that for social agents, believability arises not merely from consistency or friendliness, but from perceived autonomy, which includes the capacity for disobedience and constructive conflict. Further- more, the research shows that hybrid agents create meaningful interactions not by hiding the boundary between worlds, but by crossing it in an explicit and purposeful manner. Overall, these studies provide a comprehensive understanding of how hybrid AI agents can blend fiction and reality, introducing new dimensions and metrics for evaluating AI agents. Specifically, this research proposes two dimensions for hybrid agents: content generation based on user input and their role as actors in real-life contexts. Additionally, it identifies three key metrics for assessing hybrid agents: social expansion, spatial expansion, and language expansion. These findings contribute to the understanding of hybrid AI agents in interactive storytelling and their broader impact across gaming and real-world interactions. The research outcomes have been recognized at top-tier conferences and exhibitions, making an impact in academia, art, and industry.