RPGAgent: Driving Coherent Story-to-Play Generation with an LLM-Based Multi-Agent System

Paper Title

RPGAgent: Driving Coherent Story-to-Play Generation with an LLM-Based Multi-Agent System

Publication Info

• Topic area: AI-assisted game design using multi-agent systems for RPG prototyping.
• Keywords: RPGAgent, multi-agent systems, LLMs, procedural content generation, game design, narrative generation, scene generation, mechanics implementation, co-creative tools, Elemental Tetrad.

Background and Problem

• Problem / challenge: Existing generative AI approaches for game design struggle with integrating diverse creative elements (narrative, mechanics, aesthetics, technology) into a coherent whole. Monolithic LLMs often fail to maintain cross-domain consistency, and novice designers face significant barriers in resolving these inconsistencies during prototyping.
• Significance: Achieving coherence across game components is critical for creating immersive and believable RPGs. Addressing this challenge can lower barriers for novice designers and accelerate the prototyping process.
• Motivation and related work: Prior work has explored LLMs for narrative and visual content generation, as well as procedural content generation (PCG) for spatial design. However, these approaches often focus on isolated aspects of game design, lack holistic integration, and fail to provide structured workflows for novice creators. Multi-agent systems have shown promise in other domains but remain underexplored in game development.

Solution

• Proposed approach: RPGAgent, a multi-agent system combining LLMs and PCG techniques to transform high-level story outlines into coherent, playable RPG prototypes.
• Novelty:
  1. Embeds LLM-based multi-agent collaboration within a game engine to accelerate RPG prototyping.
  2. Implements the Elemental Tetrad framework to ensure consistency across Story, Aesthetics, Mechanics, and Technology.
  3. Combines PCG techniques with modular workflows to enhance spatial variety and structural coherence.
  4. Demonstrates a structured “story-to-play” pipeline that bridges abstract narratives and interactive prototypes.
• Procedure and key techniques:
  1. Narrative Generation: Converts user prompts into structured narrative frameworks, including world settings, character profiles, and plot steps.
  2. Scene Generation: Uses PCG techniques to create spatial environments aligned with narrative steps, employing terrain generation, asset placement, and semantic annotations.
  3. Mechanic Implementation: Defines gameplay mechanics based on narrative and spatial inputs, then generates executable code using constrained templates.
  4. System Integration: Combines a Unity front-end with a Python back-end for real-time interaction, modular agent collaboration, and iterative refinement.

Results

• Concrete findings:
  • RPGAgent significantly outperformed a GPT-assisted baseline in user experience (UEQ: 5.047 vs. 3.836, p < 0.001) and creativity support (CSI: 5.233 vs. 4.406, p < 0.001).
  • Participants rated RPGAgent higher in dimensions like efficiency, exploration, and enjoyment.
• Advantage over baselines: RPGAgent provided a seamless, integrated workflow that automated the transition from narrative to playable prototypes, reducing the need for manual intervention and tool-switching compared to the baseline.
• Experiments / evaluation:
  • Conducted a within-subjects study with 18 participants (balanced by gender and design background).
  • Compared RPGAgent to a baseline workflow using Unity Tilemap and GPT-4o for manual integration.
  • Measured user experience (UEQ) and creativity support (CSI) through surveys, complemented by qualitative interviews and interaction logs.
• Limitations and future work:
  • Framework is currently optimized for narrative-driven RPGs and may not generalize to mechanics-first genres.
  • Code generation is constrained to template instantiation, limiting flexibility.
  • Study participants were primarily students, and the evaluation was short-term; future work should include diverse user groups and longitudinal studies.

Summary

RPGAgent introduces a multi-agent system that integrates LLMs and PCG techniques to transform high-level story outlines into coherent, playable RPG prototypes. By leveraging the Elemental Tetrad framework, the system ensures consistency across narrative, aesthetics, mechanics, and technology. A user study demonstrated significant improvements in efficiency, exploration, and enjoyment compared to a GPT-assisted baseline. While the system effectively supports novice designers and accelerates prototyping, future work will focus on expanding its applicability to other game genres, enhancing code generation flexibility, and improving usability for diverse user groups.