Odyssey releases Agora-1, a multi-agent world model running up to four players in a shared GoldenEye deathmatch simulation generated entirely in real time.
Key Takeaways
Architecture decouples simulation and rendering: a state model learns gameplay dynamics from raw game internals; a DiT-based renderer generates per-player viewpoints conditioned on that shared state.
Unlike Multiverse (split-screen concatenation) and Solaris (sequence-dimension concatenation), Agora-1 maintains an explicit shared world state, avoiding context scaling problems and cross-player consistency failures.
The shared state tracks health, position, and agent metadata, and can be directly manipulated to generate new levels while preserving learned gameplay dynamics.
Integrates with PROWL, Odyssey’s RL adversarial framework, enabling multi-agent co-evolution where agents and world models push each other into harder regimes.
Team positions the architecture as domain-agnostic: collaborative robotics and open-ended foundation model training are named near-term targets beyond games.
