Agora-1: The Multi-Agent World Model

TL;DR

Agora-1 is Odyssey’s first multi-agent world model — a closed-source research preview that lets up to four players share the same generated GoldenEye deathmatch in real time, with every frame produced by the model rather than a hand-written game engine. Architecturally it splits a traditional monolithic world model into two learned components: a simulation model that evolves an explicit shared world state (positions, health, etc.) from player actions, and a DiT-based renderer conditioned on that shared state rather than on prompts/images. The decoupling buys two things existing single-stream multi-player world models (Multiverse, Solaris) struggle with — linear scaling in player count, and view consistency when players lose line of sight — and exposes the world state as an editable handle for new levels and for multi-agent RL. Released as a playable research preview; no benchmarks, parameter counts, or training-data details are disclosed.

Key claims

Multi-agent shared simulation is positioned as the next step beyond single-participant world models, with up to four players interacting in the same generated GoldenEye world in real time, all frames streamed from the model [§intro].
Two prior approaches to multi-player are named and contrasted: Multiverse concatenates agents into a “split-screen” representation (one effective world state); Solaris concatenates participants along the sequence dimension of a single AR DiT [§From Single-Agent to Multi-Agent].
Both Multiverse and Solaris are claimed to (a) scale super-linearly in compute with the number of players (Solaris’s context grows with N) and (b) struggle to maintain cross-view consistency when players lose sight of each other [§From Single-Agent to Multi-Agent].
Agora-1’s design alternative: decouple simulation from rendering. Maintain an explicit shared world state à la MultiGen, but model dynamics and rendering as two distinct learned functions [§Learning Shared World State].
The simulation model is trained directly on a game’s internal state (health, position, etc.) and predicts state transitions from player actions; this is structurally different from continuous-visual DiTs and requires bespoke architecture + structured training data [§Learning Shared World State].
The renderer is a DiT-based world model conditioned on shared game state (not on prompts or images), and must produce consistent visual outputs of the same state from multiple simultaneous viewpoints [§Learning Shared World State].
Manipulating the underlying game state directly is sufficient to generate entirely new levels while preserving the source games’ dynamics — i.e. the renderer generalizes off the state distribution it was trained on [§Learning Shared World State].
Agora-1 is framed as an RL research environment: multi-agent interaction expands the joint interaction space combinatorially with players, and passive demonstrations cover an increasingly small fraction of the meaningful behaviours (collisions, contested objectives, coordination) [§Multi-Agent Reinforcement Learning].
The same environment can serve as a generative training simulator — policies trained inside Agora-1 are speculated to transfer to unseen environments and partners without needing the original game [§Imagined Multi-Agent Training].
All technical claims are qualitative; no benchmark numbers, ablations, or architecture/parameter details are released [whole post].

Method

Agora-1 factors a multi-player world model into two learned components. A simulation model ingests player actions and updates an explicit shared game state — discrete/structured variables like position, health, and other gameplay quantities for every participant. A separate DiT-based renderer is conditioned directly on this shared state (not on text prompts or reference images) and produces each player’s first-person view of the same world. Because the renderer’s conditioning is the same authoritative state object for all participants, generating N viewpoints reduces to running N rendering passes against one state, and cross-view inconsistencies (player A sees an object that player B can’t, players walking away from each other and back) reduce to whether the renderer is faithful to its state input. Both components are entirely learned: the simulation model is trained on the source game’s internal state stream rather than on pixels, and the renderer is trained jointly to map state → multiple viewpoints. The post draws an analogy to a modern game engine’s logic/render split but emphasizes that nothing here is hand-coded. The architecture is illustrated as a single state stream feeding parallel rendering passes (one per player) — a structural contrast to Solaris’s single AR DiT that absorbs all participants into one sequence.

Concrete deployment: GoldenEye-style 4-player deathmatch, real time, frames streamed to every participant simultaneously. The post says nothing about model size, training compute, FPS, or context length.

Results

No quantitative results, ablations, or benchmark numbers are released. The published artifact is a playable research preview plus an architecture description and demo videos of (a) a 4-player shared deathmatch and (b) PROWL-style adversarial-policy exploration of the world model. All comparative claims against Multiverse and Solaris are made as architectural arguments, not empirical comparisons.

Why it’s interesting

Agora-1 is the cleanest architectural counter-design to Solaris that the wiki has filed so far for multi-agent video world models, and it converges with MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines on the “hoist shared state out of the network” pattern — but applies it inside a closed-flagship stack rather than as an open-research paper. Reading MultiGen and Agora-1 together pins the design point: external authoritative state + per-player rendering passes is now the design alternative to in-context multi-agent concatenation, in both the open-research and closed-flagship lanes. The post also extends Odyssey’s previously single-player flagship line (Introducing Odyssey-2 Max: Scaled World Simulation) with a multi-player axis, suggesting the closed labs are starting to compete on interactivity primitives (single-player long-horizon vs. multi-player shared state) rather than only on physics-benchmark numbers. The framing of Agora-1 as both an interactive product and an RL environment for adversarial-policy training (PROWL) makes it the first filed datapoint where a closed-flagship WFM is presented as a training substrate for its own data flywheel, which sharpens the “WFM as backbone” vs “WFM as environment” axis already raised in World Foundation Models.