Agora-1: Open Source Multi-Agent World Model for Real-Time Shared Simulation

Odyssey has released Agora-1, the first multi-agent world model that allows multiple participants—human or AI—to share and interact within the same generated world simulation in real time. The model supports up to four players in a shared deathmatch simulation, where every pixel is generated by the model in real time, functioning as a learned game engine.

Architecture: Decoupled Simulation and Rendering

Agora-1 separates the world model into two distinct learned components:

• Simulation model: Trained on the internal game state (e.g., GoldenEye), learning gameplay dynamics and how state transitions occur from player actions.
• Rendering model: A DiT-based world model conditioned on the shared game state (not prompts or images) to generate consistent visual representations from multiple viewpoints simultaneously.

This decoupling is analogous to a modern game engine, but both components are entirely learned from data. The model manipulates the underlying game state directly, enabling generation of entirely new levels while preserving gameplay dynamics.

Key Capabilities

• Up to 4 concurrent participants in a shared simulation.
• Real-time pixel streaming generated by Agora-1.
• Shared world state tracks health, position, and other agent attributes.
• Can generate novel levels consistent with the source game's dynamics.

Comparison to Previous Work

Prior approaches like Multiverse concatenate agent states into a single representation, while Solaris stacks participants along the sequence dimension (not scaling linearly with player count). Both struggle with consistency when players lose sight of each other. Agora-1's decoupled approach avoids these limitations.

Use Cases

Odyssey targets applications in gaming, robotics, defense, education, and foundation model training. The architecture can scale to handle increasingly complex simulations and state representations beyond GoldenEye.