Up to 20 humans and agents share the same World Model with Agora-2

Agora-2 expands Odyssey's multi-agent World Model into shared environments supporting up to 20 humans and agents in real time. Every action modifies a common state that the model then uses to generate each participant's individual perspective.

One environment for 20 participants

An interactive World Model typically generates one user's experience based on their own actions. With Agora-2, Odyssey is maintaining multiple participants inside the same simulation, where each person's or agent's actions can affect what everyone else experiences next.

The system supports up to 20 humans and agents, five times more participants than Agora-1 according to Odyssey. The currently available multiplayer research preview uses a configuration of up to four humans facing sixteen agents.

To explore this architecture, Odyssey trained Agora-2 on Diablo II captures pairing observations with actions and state. The World Model learns how multiple entities move and interact without running the original game engine during simulation. A shared state instead of a sequence of images

Agora-2 separates simulation from each participant's individual rendering. A simulation model first predicts how their combined actions modify the shared state based on entity properties, recent actions, and surrounding geometry.

A world server combines those predictions to maintain that common state. An entity that leaves someone's field of view therefore continues to exist in the internal representation instead of having to be reconstructed solely from visual history when it reappears.

Each participant then receives a separate view of the simulation. The rendering model uses the shared environment, participating entities, and recent visual history to generate a different perspective for every human or agent in real time. Agents trained to pursue and cooperate

Agents inside Agora-2 are trained through reinforcement learning to pursue opponents, navigate around obstacles, and recover when they become stuck or separated.

They do not receive a complete representation of the environment. Decisions are made from a partially observed view and recent observations, requiring agents to track nearby participants and adapt their behavior as the simulation changes.

Odyssey connects this work to PROWL, its research into agents that explore simulated environments and generate new training data through their interactions and failures. Letting agents and environments learn together

The longer-term direction is to bring this multi-agent architecture into more general World Models such as Odyssey-3. The research aims to move beyond environments learned from a specific game toward representations capable of handling unfamiliar entities, actions, and relationships.

Odyssey points to robotics, autonomous vehicles, cybersecurity, AI training, and gaming as potential applications. These remain research directions described by the company rather than capabilities demonstrated by Agora-2.

The Agora-2 research preview is currently available to test the multiplayer environment directly, alongside a separate technical report covering its architecture and simulation system.