Guiding, Not Railroading: Design and Evaluation of a Multi-Agent System for Narrative Redirection in Role-playing Games

Large Language Models (LLMs) are poised to revolutionize interactive storytelling, yet they introduce a fundamental HCI challenge: balancing player agency with the coherence of a pre-authored narrative. Existing LLM-driven Game Masters (GMs) often undermine the player experience by being overly compliant, leading to disjointed stories, or by rigidly "railroading" players, which diminishes their sense of freedom. This paper addresses this tension by introducing SENNA, a novel multi-agent AI system designed to maintain narrative adherence in single-player role-playing games. SENNA operationalizes a prewritten adventure by creating and using a Narrative Graph to model the structure of the story, track the player's progress and ensure key plot points are met while maintaining interactivity. A primary contribution of this work is an empirical investigation of narrative redirection, the methods an AI GM can use to guide players back to the intended arc when they deviate. We designed six redirection strategies inspired by expert human GMs and conducted a user study to evaluate their effectiveness. The study combined live gameplay with a within-subjects comparison of different strategies at key narrative junctions. Our findings show that players strongly prefer redirection techniques grounded in the internal logic of the game, such as offering more information, experiencing consequences in the world, and influence from non-player characters. These methods successfully maintained narrative coherence without sacrificing perceived autonomy or immersion, while simplistic "hard denials" were poorly received. Our work contributes an empirically validated framework and actionable design guidelines to create more robust and human-centered AI-driven narrative experiences.