Following the success of the Concordia Contest held by the Cooperative AI Foundation and collaborators at NeurIPS 2024, we are excited to announce the release of an updated Concordia library by Google DeepMind. v2.0 expands on the original framework, providing a more robust environment for creating multi-agent simulations, games, and AI evaluations.
We have invited Alexander Sasha Vezhnevets and Joel Z. Leibo, primary architects behind Concordia and senior contributors to our recent competition, to take us through key new features or improvements in the new version and how it can be used.
Concordia is a library for generative agent-based modelling that leverages generative AI to create multi-actor environments for a wide range of purposes, from social science modelling to interactive storytelling.
Inspired by tabletop role-playing games (TTRPGs), Concordia v2.0 uses an Entity-Component architectural pattern, which makes it incredibly flexible and user-friendly. This design allows users to act as ‘designers’ rapidly building and configuring complex scenarios by mixing and matching reusable components without having to write new code.
Whether you're an evaluationist who needs a rigorous testbed for AI, a dramatist who wants to generate compelling narratives, or the simulationist who seeks to model and understand real-world phenomena, Concordia provides the tools to create environments tailored to your specific goals.
In Concordia v2.0 we focused on making the system simpler, more unified and lightweight: game masters are now entities, and clocks are optional. Interactions with memory have been simplified. New concepts include "prefabs" replacing "factories," and an "engine" to structure agent-game master interaction. We also added support for easy serialisation and checkpointing.
Explore the tech report for more details: https://arxiv.org/pdf/2507.08892
And the library itself here: https://github.com/google-deepmind/concordia
To find out more, make sure to watch our recently released tutorial on Concordia v2.0's capabilities - including what it's for, how it's structured and how you can start working with it to design evaluations.
The Cooperative AI Foundation's paper based on the Concordia Contest is currently under review at NeurIPS. We'll be releasing outputs on this in the near future. Make sure you're subscribed to our mailing list to stay informed.
August 21, 2025
.png)
