Five LLM agents play Werewolf in-browser, each with a private DuckDB (kayhan.dev)

A recent demonstration showcased how five language model (LLM) agents play the game Werewolf directly in a browser, each operating with a private DuckDB-WASM database contained within its own Web Worker. The significance of this development lies in its innovation of enforcing information asymmetry at the database level rather than through application code. This architecture allows each agent to possess a unique schema for managing its data—such as intentions, knowledge, and actions—while a centralized gateway worker controls data flow using scoped permissions. The agents can freely interact in gameplay while safeguarding their private reasoning from other players. Technically, each Web Worker maintains its own DuckDB-WASM instance, contributing to the overall game state through controlled and token-authenticated database interactions. The orchestration of gameplay is managed by a single entity that cycles through each agent, calling on their private data to drive decisions while preventing unauthorized access to sensitive information. This layer-based architecture exemplifies a robust model for developing multi-agent systems, where data access can be finely tuned, allowing for secure and isolated decision-making while maintaining a cohesive gameplay experience. This approach presents exciting implications for future developments in AI systems that require strict information management among disparate agents.