SAN FRANCISCO: AI research company Odyssey has introduced Agora-2, a next-generation multi-agent world model designed to allow humans and artificial intelligence agents to interact inside the same virtual environment in real time.

The system represents a new direction for world-model research, moving beyond AI systems that simulate an environment for a single participant. Agora-2 can maintain a shared simulated world in which the actions of multiple participants influence what happens to everyone else.

According to Odyssey, Agora-2 can support up to 20 humans and AI agents simultaneously. A playable research preview allows multiple participants to enter the same generated environment, including scenarios involving human players and AI-controlled agents.

Rather than relying on a conventional game engine to generate the environment, Agora-2 uses AI models to simulate the world and produce the visual experience. This allows the system to predict how participants’ actions affect the shared state of the environment.

The technology builds on Odyssey’s earlier Agora-1 system but significantly increases the number of participants that can interact within the same simulated world. The company says Agora-2 supports five times more participants than its predecessor.

For its research preview, Odyssey trained the system using captured gameplay from Diablo II. The training data combines observations, actions and information about the state of the environment, allowing the model to learn how entities move, interact and respond to changes around them.

One important feature is Agora-2’s shared state system. Instead of independently generating what each participant sees, the system maintains information about entities and their interactions within a common simulated state. Individual views are then generated from that shared environment.

This approach is designed to keep participants and objects consistent even when they temporarily move outside another participant’s field of view.

The technology could eventually have applications beyond gaming. Odyssey says multi-agent world models could be useful for training AI systems in areas such as robotics, autonomous vehicles, cybersecurity, energy and other environments where several independent agents must cooperate or compete.

The approach could also provide researchers with a controlled environment for studying how AI agents behave when they interact with other intelligent systems.

For example, multiple AI agents could be trained to coordinate tasks, compete against one another or respond to changing conditions without requiring researchers to expose experimental systems to real-world environments.

Odyssey is also exploring reinforcement-learning techniques in which AI agents learn to navigate environments, pursue objectives, overcome obstacles and recover from difficult situations.

The company describes Agora-2 as a research preview rather than a finished commercial platform. Its current demonstrations are focused on a particular type of simulated environment, meaning significant research will still be required before the technology can be generalized to a wide range of real-world scenarios.

Long-term challenges include maintaining consistency during complex interactions, supporting larger numbers of participants and extending the system to unfamiliar environments and objects.

Despite those limitations, Agora-2 demonstrates how AI-generated environments could evolve from simple visual demonstrations into interactive simulations where multiple humans and autonomous agents share the same virtual world.

The development could become particularly important for AI training, because simulated environments can provide agents with opportunities to learn through repeated interaction while reducing the risks and costs associated with conducting certain experiments in the physical world.