Towards LLM-Agents that Play Dungeons Dragons Using Iterative Prompting

Playing Dungeons & Dragons (D&D) requires players to attend every session of a campaign, which is often hindered by scheduling conflicts. The disruptions caused by these conflicts could lead to stalled or cancelled games. To address this challenge, we propose using AI player substitutes. Specifically, we develop LLM-agents capable of social interaction and exploration within textual D&D scenarios, using Concordia—a library to simulate multi-agent dialogue and interactions. We iteratively prompt our LLM-agents to foster collaboration, task progression, and comply with the narrative context in two distinct campaign settings. We annotate natural language transcripts generated in Concordia, categorising player actions into the three action and dialogue categories. Our preliminary findings indicate that iterative prompting enhances agents’ narrative compliance, collaborative behaviour and progression towards campaign goals. These promising results suggest LLM-agents could be viable stand-ins for human players, and warrant further investigation.