Multi-Actor Generative Artificial Intelligence as a Game Engine

Alexander Sasha Vezhnevets, Jayd Matyas, Logan Cross, Davide Paglieri, Minsuk Chang, William A. Cunningham, Simon Osindero, William S. Isaac, Joel Z. Leibo

Abstract

Generative AI can be used in multi-actor environments with purposes ranging from social science modeling to interactive narrative and AI evaluation. Supporting this diversity of use cases—which we classify as Simulationist, Dramatist, and Evaluationist—demands a flexible scenario definition framework. We argue here that a good approach is to take inspiration from tabletop role-playing games (TTRPGs), where a Game Master (GM) is responsible for the environment and generates all parts of the story not directly determined by the voluntary actions of player characters. We argue that the Entity-Component architectural pattern is useful here. In such a system, the GM is not a hardcoded computer game but is itself a configurable entity, composed of components just like any other actor. By design, the approach allows for a separation between the underlying implementation details handled by an engineer, the creation of reusable components, and their composition and configuration managed by a designer who constructs entities from the components. This separation of concerns is instrumental for achieving rapid iteration, maintaining modularity, and ultimately to ensure scalability. We describe the ongoing evolution of the Concordia library in terms of this philosophy, demonstrating how it allows users to effectively configure scenarios that align with their specific goals.

Introduction

Building "agentic" multi-actor environments—where multiple AI agents interact with each other and a simulated world—has historically been a fragmented and brittle process. Researchers often find themselves hardcoding environment rules and agent behaviors for every new experiment, leading to significant technical debt and limited reusability. Furthermore, the field is pulled in different directions by conflicting goals: social scientists seek high-fidelity simulations, storytellers prioritize dramatic engagement, and AI researchers require standardized benchmarks.

To address these challenges, researchers from Google DeepMind and the University of Toronto propose a new architectural paradigm: treating multi-agent generative AI systems like a game engine. By adopting the Entity-Component (EC) pattern common in modern game development, they have developed a framework that separates low-level engineering from high-level scenario design.

The Entity-Component Architecture

The core of this proposal is the transition from rigid class hierarchies to a composition-based model. In the Entity-Component pattern:

  • Entities: These are generic containers representing any actor in the system, such as an agent, a corporation, or a legal body.
  • Components: These define specific behaviors or states (e.g., memory, social obligations, or planning capabilities).

By attaching different components to an entity, designers can create complex behaviors without writing new low-level code. This modularity allows for "prefabs"—pre-configured collections of components—that enable the rapid deployment of complex simulations, such as a populated town or an economic market.

The Role of the Generative Game Master

In traditional simulations, the environment is a set of hardcoded rules. This framework replaces that rigidity with a Generative Game Master (GM). Inspired by tabletop role-playing games (TTRPGs), the GM is itself a configurable Entity managed by a Large Language Model (LLM). The GM is responsible for everything outside the players' direct control: describing the world, mediating interactions, and enforcing consequences.

This approach allows researchers to swap GM "personalities" depending on the experiment. For instance, a "Narrative Director" GM might prioritize dramatic tension, while a "Strict Evaluator" GM might focus on rigorous rule adherence.

A Typology of Research Motivations

The paper identifies three primary motivations for building multi-actor systems, mapping them to TTRPG player types:

  1. Simulationist (The Scientist): Focused on high-fidelity modeling of real-world behavior and causal physics.
  2. Dramatist (The Storyteller): Focused on creating engaging, coherent, and aesthetically pleasing narratives.
  3. Evaluationist (The Judge): Focused on fair, standardized benchmarks to measure model performance and social intelligence.

Recognizing these distinct goals is crucial, as they often involve design trade-offs. A GM optimized for drama might "fudge" rules to keep a story moving, which would compromise a simulationist study of economic behavior.

The Concordia Library

The authors implement this philosophy through Concordia, a library designed for rapid iteration in multi-agent research. Concordia facilitates a clear separation of concerns:

  • Engineers focus on building reusable, testable Python/LLM components.
  • Designers (researchers and users) assemble these components into specific scenarios.

Concordia supports diverse interaction dynamics, including simultaneous turns (ideal for market trading) and sequential turns (ideal for dialogue). It also treats synthetic data generation as a byproduct of the simulation process rather than a separate task, allowing researchers to generate high-quality reasoning or creative data by tuning the simulation's motivation. This utility has already been demonstrated in projects like the Concordia Contest at NeurIPS 2024.

Conclusion and Future Work

The game engine paradigm offers a scalable and flexible path forward for multi-actor AI. By decoupling cognitive architectures from environmental backdrops, researchers can more effectively evaluate different LLMs in complex social contexts.

While the system currently faces challenges related to the computational cost of LLM calls and the "steerability" of models, the move toward a "Multi-Actor System Definition Language" promises to further simplify the design of sophisticated AI environments. The ongoing development of the Concordia library provides a practical foundation for this evolving field.

Summary updated

Abstract

Generative AI can be used in multi-actor environments with purposes ranging from social science modeling to interactive narrative and AI evaluation. Supporting this diversity of use cases -- which we classify as Simulationist, Dramatist, and Evaluationist -- demands a flexible scenario definition framework. We argue here that a good approach is to take inspiration from tabletop role-playing games (TTRPGs), where a Game Master (GM) is responsible for the environment and generates all parts of the story not directly determined by the voluntary actions of player characters. We argue that the Entity-Component architectural pattern is useful here. In such a system, the GM is not a hardcoded computer game but is itself a configurable entity, composed of components just like any other actor. By design, the approach allows for a separation between the underlying implementation details handled by an engineer, the creation of reusable components, and their composition and configuration managed by a designer who constructs entities from the components. This separation of concerns is instrumental for achieving rapid iteration, maintaining modularity, and ultimately to ensure scalability. We describe the ongoing evolution of the Concordia library in terms of this philosophy, demonstrating how it allows users to effectively configure scenarios that align with their specific goals.
BibTeX
@misc{vezhnevets2025multi,
  author = {Alexander Sasha Vezhnevets and Jayd Matyas and Logan Cross and Davide Paglieri and Minsuk Chang and William A. Cunningham and Simon Osindero and William S. Isaac and Joel Z. Leibo},
  title = {{Multi-Actor Generative Artificial Intelligence as a Game Engine}},
  year = {2025},
  eprint = {2507.08892},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.08892},
}
Figure