Odyssey lets 20 humans and AI agents play together in a generated world

Most world models we have seen so far feel like single-player games. The neural network generates pixels exclusively around you, and the entire virtual reality revolves around your lone point of view.

Odyssey released a playable research preview of Agora-2, its next-generation multi-agent world model. It takes world simulation from solo play into full multiplayer, supporting up to 20 humans and AI agents interacting simultaneously inside the same generated world—a 5x jump over Agora-1.

A shared brain behind 20 screens

The hardest problem with putting 20 actors inside a generative world is consistency. If every screen hallucinates its own reality, the simulation immediately falls apart.

Odyssey solved this by acting like a learned game engine. Instead of generating pixels directly from raw inputs, Agora-2 splits the work. A simulation model takes in the properties, geometry, and recent actions of all 20 participants, predicts how they affect one another, and updates a single shared global state on a world server.

Because this state exists independently of any single camera, an agent doesn't vanish or break causality when it walks off your screen. A rendering model trained with flow matching then takes that global state and streams custom, coherent viewpoints to all 20 participants in real time.

Learning physics from Diablo II

To teach the model how multiple entities coordinate and collide, Odyssey turned to Blizzard's classic Diablo II.

Rather than coding movement rules and collision boundaries by hand, they trained Agora-2 on vast gameplay captures pairing visual observations with actions and game states. The model essentially reverse-engineered navigation, line of sight, and combat directly from data.

Alongside the world model, Odyssey trained agents with reinforcement learning to hunt targets, path around obstacles, and recover when separated. Through a project called PROWL, the team is now using the edge cases and failures uncovered during these interactions to train and improve the world model itself.

Why it matters

We think Agora-2 represents a genuine shift in how synthetic training grounds will be built.

Until now, training teams of autonomous vehicles, warehouse robots, or defensive cyber agents required massive teams of engineers to hardcode rules into rigid physics simulators. Agora-2 shows that a neural network can learn to host multi-agent environments from scratch, maintaining consistency across dozens of actors without human-written game engines.

It is especially urgent for agent security. As researchers confront emerging risks like multi-agent collusion and coordinated cyberattacks, safe sandboxes to study adversarial behavior become critical. Agora-2 lets researchers run those chaotic scenarios without exposing any real-world infrastructure.

That said, running real-time generative simulations for 20 streams remains computationally heavy. While the playable preview proves the concept works, scaling this architecture to hundreds of participants over extended hours without hallucination drift will be the next major engineering hurdle.