Odyssey's Agora-1 turns GoldenEye into a learned multiplayer game engine

Odyssey has unveiled Agora-1, a world model that generates a shared simulated environment for up to four simultaneous participants, demonstrated as a real-time deathmatch built on GoldenEye. Unlike prior multi-agent attempts such as Multiverse and Solaris, which either glue player views together or extend a single autoregressive transformer’s context, Agora-1 splits the problem into two learned components: one model tracks the evolving game state (positions, health, actions) and a separate DiT-based renderer produces consistent pixels for each player’s viewpoint from that shared state.

The decoupling mirrors the structure of a conventional game engine but replaces hand-coded physics and rendering with learned systems trained on internal game state. Because the state representation is explicit, the system can be manipulated to generate novel levels while preserving the source game’s dynamics, and it sidesteps the consistency failures that plague competing approaches when players lose line of sight.

Odyssey frames the work less as a gaming product than as a substrate for multi-agent reinforcement learning and foundation-model research. With combinatorial interaction spaces that passive datasets cannot cover, the company argues that learned multi-agent simulators are needed to generate training experiences for more general agents, with downstream relevance to collaborative robotics and other domains requiring shared simulated environments.