Concordia

Concordia is a library for what its authors call generative agent-based models: simulations of many actors, each driven by a large language model (LLM), that act in a shared world.^1,2 It lives in Google DeepMind's GitHub organisation, and the README states that it is not an officially supported Google product.¹ The 2023 technical report has ten authors, led by Alexander Sasha Vezhnevets, with Joel Z. Leibo last.^1,2

The design borrows from tabletop role-playing games.^1,2 Characters say in plain language what they intend to do. A special entity, the game master, decides what happens, for instance by checking whether an action is physically possible, and tells each character what it observes.^1,2 In a simulation of digital services the game master can also call real apps and APIs.² Each entity is assembled from components that hold its instructions, memory and way of choosing actions, and ready-made templates called prefabs give a starting point.¹ The README's worked example is four friends snowed in at a pub, two of them arguing over a crashed car.¹

You install it with pip and write Python, or use a graphical editor and command-line interface that come with it.¹ It needs any LLM API that can generate text, plus a text embedder for searching memory. Results depend on the model chosen.¹ Tutorials include a notebook that runs in Colab and a recorded lecture.¹ The authors list social science, AI safety and ethics, economics, synthetic data and testing real services with simulated users among its uses.^1,2 Version 2.4.0 came out in March 2026,³ and commits continued into October 2026.⁴

Concordia is a toolkit, not a model of any population. It has no census data, map or calibrated behaviour built in. Whoever builds a simulation with it has to show that the characters stand for real people. Licence Apache-2.0.¹