Generating Board Games with MCTS and LLM Agents

Abstract

Large Language Models (LLMs) have shown significant potential in Procedural Content Generation (PCG). Yet, their application to the generation and balancing of interactive game mechanics remains largely unexplored. While PCG is a mature field, most research focuses on static assets rather than executable mechanics. Existing systems for rule synthesis frequently utilize domain-specific languages, e.g., the Ludii Game Description Language (Ludii GDL), which impose structural constraints. By contrast, generating games in Python leverages the extensive pre-training data of LLMs. To address this gap, we introduce an agentic pipeline for the procedural generation and dynamic balancing of board game mechanics. From a natural language prompt, a programmer LLM synthesizes executable rulesets. An MCTS agent then runs playouts to extract viability metrics, such as completion, balance, agency, and coverage. These metrics are compared against classic strategy baselines, and a Mutation Agent uses the largest deviation to guide targeted rule or code mutations. Our evaluation demonstrates that this metric-informed, closed-loop approach effectively optimizes the procedural search space, allowing the system to iteratively refine and balance playable game mechanics.