DeepMind Empowers AI as 'God', Directing a 'Westworld'-like World of AI Actors}

Have you played Werewolf or other social deduction games? These are classic tabletop role-playing games (TTRPGs), where the Game Master (GM) acts as the director, writer, and narrator, controlling the environment, storytelling, and NPCs.

Now, imagine replacing the GM with a powerful generative AI, while the players are also AI agents with distinct minds. What kind of world would this create?

This setup could enable:

• Scientific Simulation: Building virtual societies for social science research and observing emergent group behaviors.
• Interactive Narratives: Creating dynamic stories or games where AI agents role-play and co-create plots.
• AI Evaluation: Designing scenarios as 'examinations' to test AI capabilities like reasoning, collaboration, and communication.

      ^{The future Westworld, where all characters are AI in a virtual Western-themed world.}

However, these needs—scientific, dramatic, and fair—are vastly different and sometimes conflicting. How can a unified framework satisfy all?

Researchers from DeepMind and the University of Toronto drew inspiration from TTRPGs and modern game engines, proposing a solution: a software library called Concordia.

• Paper: Multi-Actor Generative Artificial Intelligence as a Game Engine
• Project: https://github.com/google-deepmind/concordia

Traditionally, game environments are hardcoded. The idea here is to make the GM itself a configurable, AI-driven agent.

Concordia’s design is based on the modern game engine architecture of Entity-Component. In this framework, both AI players and the GM are just entity containers, with capabilities defined by pluggable components like memory, goals, or social rules.

This approach separates the roles of engineers and designers: engineers create powerful components, while designers assemble these like LEGO to quickly build and test complex scenarios, mostly without coding.

Entities, Components, and Game Design

The entity-component architecture provides a flexible foundation for multi-agent generative AI systems.

Entities are lightweight objects with unique IDs, with behaviors and attributes determined by attached components. The engine calls functions like observe and act, implemented by components.

Components combine Python code and LLM calls, offering maximum flexibility. Designers can implement specific functions or let the GM narrative LLM generate content, often coexisting in the same environment.

Entities support two main calls: observe and act. Components implement preobserve, postobserve, preact, and postact methods, enabling rapid creation of diverse entities by component combination—much more flexible than traditional inheritance models.

For generative AI agents, this architecture allows multiple components—Memory, Planning, Beliefs—to form a cognitive system. Similarly, organizational entities can be built from components representing departments, policies, and communication structures.

The GM itself is an entity, customizable via components, enabling flexible roles such as strict evaluator, narrative guide, or causal maintainer, depending on the multi-agent system’s needs.

Concordia also supports various game engine modes for different interaction dynamics.

Design Goals for Game/Simulation

According to Edwards, a key figure in tabletop RPG theory, TTRPGs can be classified as:

• Gamist: GM designs challenges to maintain fun.
• Narrativist: GM adjusts story to respond to player input.
• Simulationist: Players seek immersion in a logically consistent virtual world.

Using multi-agent generative AI, motivations can be categorized as:

• Evaluationist: Focused on fairness and performance comparison.
• Dramatist: Focused on storytelling and character development.
• Simulationist: Focused on realistic world-building.

AI can also generate synthetic training data for these scenarios.

Evaluation Perspective

Evaluation users seek to compare AI systems across metrics in a fair environment with clear success indicators, such as standardized scenarios and performance benchmarks.

Drama Perspective

Drama users prioritize narrative coherence, emotional resonance, and dynamic character interactions, emphasizing storytelling over strict evaluation.

In future sections, the paper discusses simulation, synthetic data, and other research areas. Interested readers can refer to the original paper for more details.