← Back to feed

Odyssey released Agora-1 on May 18 — the first world model architected for multiple simultaneous participants. Until now, world models have been single-player: one agent, one stream of actions, one generated perspective. Agora-1 changes that.

What It Does

Up to four players, human or AI, interact inside the same generated world at once. Every participant sees a consistent view of the same simulation, generated in real time. Odyssey’s research preview uses a GoldenEye deathmatch: four players navigate the same rendered level, with the model tracking interactions, collisions, and consequences across all four action streams simultaneously.

The key constraint Agora-1 solves: shared world state consistency. A single-player world model can drift slightly, or hallucinate, because there’s only one observer. With multiple agents acting from different positions, the model must maintain a coherent underlying reality that all views derive from — otherwise players see contradictory states.

The Architecture

Agora-1 separates two functions that earlier world models combined:

1. Simulation dynamics — A model trained directly on GoldenEye’s internal game state. It learns how state transitions occur from player actions: where agents move, what they can see, how the environment changes. This model understands the rules without needing to render anything.

2. Rendering — A DiT-based model conditioned on the shared game state. Given the current world state and a player’s position/orientation, it generates that player’s visual perspective. Multiple calls to the renderer — one per player viewpoint — all draw from the same consistent underlying state.

This decoupling is what makes consistency tractable. The simulation model maintains truth; the renderer produces views of that truth. Previous approaches tried to learn both jointly, which made multi-player consistency extremely hard.

Why It Matters for Agent Training

Agora-1’s most interesting application may be training policies entirely inside generated worlds. Multi-agent RL typically requires either a real environment or a purpose-built simulator. Agora-1 is a learned simulator — one where multiple agents can interact, compete, cooperate, and learn without access to the original game or environment.

Odyssey points to this explicitly: policies trained inside Agora-1’s generated worlds may generalize to unseen environments and unseen interaction partners, because the world model provides variation beyond what any single real deployment would supply.

Combined with Odyssey’s prior work on PROWL — where world models improve through active agent exploration — the architecture suggests a path toward self-improving simulation environments.

Current Limitations

Agora-1 is a research preview. The demo is limited to GoldenEye. The model is trained on that game’s internal state specifically, not a general-purpose simulator. Generalization to other games or real-world environments is unproven.

Odyssey has a public playable demo available at odyssey.ml.