Odyssey presents Agora-1, its first multi-agent world model

Odyssey releases Agora-1, the first multi-agent world model simulating a shared GoldenEye game environment in real-time for up to four players.

Odyssey Lab releases Agora-1, presented as the first world model capable of simulating a shared environment among multiple participants, human or AI, in real-time. For this first iteration, the team relied on GoldenEye, a first-person shooter that several of its members played as children. Up to four players can compete in a deathmatch entirely generated by the model, which calculates interactions, maintains a consistent world state, and renders pixels to each simultaneously.

The approach breaks with those of Multiverse or Solaris by dissociating simulation and rendering. A first network learns the game's dynamics from its real internal state: positions, health points, character actions. A second, based on a Diffusion Transformer, translates this shared state into coherent images, seen from multiple independent angles. The system thus functions as a learned game engine, where logic and rendering are no longer hand-coded but entirely deduced from data.

Odyssey claims several applications beyond gaming: reinforcement learning agent training in open environments, collaborative robotics, multi-view simulations. Agora-1 is part of the lab's recent work around Odyssey-2 and PROWL, an adversarial framework that trains an agent to uncover a world model's flaws to generate new training data.

A public demo is available for testing. This is a research preview.