Agora-1 - Odyssey's First Multi-Agent World Model

What is Agora-1

Agora-1 is Odyssey 's first multi-agent world model, supporting interaction between humans and AI in the same real-time generated world simulation. The model uses the classic FPS game GoldenEye 007 as its research scenario, supporting up to 4 participants sharing a deathmatch experience. Unlike traditional world models, Agora-1 maintains an explicit shared world state by decoupling simulation dynamics from visual rendering, achieving multi-view consistent generation. The model architecture is similar to a modern game engine, where all components are learning systems without hardcoded logic, and can be used in fields such as gaming, robotic collaboration, reinforcement learning training, and foundation model research.

Agora-1's Main Features

• Multi-Agent Real-Time Interaction: Supports up to 4 human or AI participants sharing and interacting in real time within the same generated world.
• Shared World State Maintenance: Through an explicit world state database, ensures all participants see a consistent simulated environment.
• Decoupled Simulation and Rendering: Separates the dynamic simulation and visual generation modules, supporting consistent image generation from multiple independent viewpoints.
• Learned Game Engine: Fully based on data learning, without the hardcoded logic or rendering rules of traditional game engines.
• Playable Demo Experience: Uses GoldenEye deathmatch as the scenario, providing a directly playable web-based research preview.

Agora-1's Technical Principles

• Dual-Model Decoupled Architecture: Adopts an independent Simulation Model and Rendering Model, connected through a shared World State.
• Game State Learning: The Simulation Model directly learns the internal state transition rules of the game, mastering gameplay dynamics and the impact of player behavior.
• DiT Conditional Rendering: The Rendering Model is based on the DiT architecture, directly generating visual imagery conditioned on the shared game state, rather than traditional prompts or images.
• Explicit State Management: Maintains a discrete world state containing information such as player positions and health values, supporting direct manipulation to generate new levels.
• Multi-View Consistency: Through shared state rather than concatenated context, solves the problem of maintaining consistency when multiple players' views are separated.

How to Use Agora-1

• Visit the Experience Page: Visit the Agora-1 project official website https://odyssey.ml/introducing-agora-1。
• Create a Character Identity: Enter a custom player name and enter the matchmaking waiting room.
• Wait for Matchmaking to Start: The system supports starting with 2-4 people; wait for other players to join or select "Start Anyway" to begin directly.
• Master the Controls: Use WASD to move, left and right arrow keys to adjust the view, and the spacebar to fire in the deathmatch.
• View Match Rankings: After the match ends, view kill, death, and score statistics, and choose to restart or share with friends.

Agora-1's Core Advantages

• Strong Multi-View Consistency: Generates multi-view imagery from the same shared state, avoiding visual corruption caused by separated player views.
• High Scalability: The internal state representation can be arbitrarily extended, supporting more complex simulations and gameplay dynamics.
• Linear scalability: Unlike serial sequence methods, increasing the number of participants does not cause the context length to explode.
• Direct state manipulation: The underlying game state can be modified to directly generate new levels while maintaining consistency with the source gameplay.
• Open-ended training: Supports multi-agent reinforcement learning, generating emergent behavior data through interaction that traditional demonstrations cannot cover.

Agora-1 project address

• Project official website：https://odyssey.ml/introducing-agora-1
• Online demo：https://agora.odyssey.ml/

Comparison of Agora-1 with similar competitors

Application scenarios of Agora-1

• Multiplayer game development: Provides a real-time world generation solution for AI-native multiplayer games without the need for traditional engines.
• Collaborative robots: Supports multiple robots jointly reasoning about actions, space, and interactions in a shared environment.
• Reinforcement learning research: Provides a multi-agent RL training environment, generating emergent interaction data such as collisions and coordinated movement.
• Foundation model training: As a generative multi-agent simulator, trains policies that can generalize to new environments and new partners.
• Defense and education simulation: Builds simulation environments for complex multi-participant scenarios for training and research.