Odyssey opens Agora‑2 preview, says 20 can share one generated world
The research lab says its multi-agent world model generates a shared environment for up to 20 humans and agents at once, with no game engine underneath.
Odyssey has opened a playable research preview of Agora-2, a world model the lab says lets up to 20 human and agent participants act inside the same generated environment in real time. The preview drew a run of posts on Thursday afternoon, three days after Odyssey published the model's page on 21 September.
Odyssey describes Agora-2 as "a learned game engine", and says it holds an explicit shared state across participants rather than rendering each player a separate video. Every claim about how it works is the lab's own. Odyssey has published no benchmark, no weights and no paper, and no outside group has tested the preview.
The lab puts the capacity at five times that of Agora-1, its previous model, and says Agora-2 can simulate multiple environments rather than one, with longer-horizon behaviour. It says the model uses attention to predict how participants' combined actions change the shared state, and flow matching to train the renderer that draws each participant's view.
The account Documenting AGI, which had 368 likes and 27,464 views on its post by 18:00 UTC on Thursday, said the preview generates shared environments across multiple worlds with no traditional game engine under the hood. The account 0xHvdes said the generated world looks like Diablo II. Both are describing the demo rather than a controlled test.
Odyssey says it intends the worlds for training robots, autonomous vehicles and cybersecurity agents, and for studying how agents behave together before they are deployed. The preview is at agora.odyssey.ml, which redirects to the lab's own domain.