# Agora-1

> Source: https://aiwiki.ai/wiki/agora_1
> Updated: 2026-09-16
> Fact-checked: 2026-09-16
> Categories: AI Models, AI in Gaming, Reinforcement Learning, World Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Agora-1." aiwiki.ai, 16 Sept 2026. https://aiwiki.ai/wiki/agora_1
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Agora-1 is a multi-agent [world model](https://aiwiki.ai/wiki/world_model) released by the AI lab [Odyssey](https://aiwiki.ai/wiki/odyssey_ai) on 18 May 2026. It generates one shared simulation that up to four participants, human or AI, can occupy and act in simultaneously, with each participant receiving a separately rendered viewpoint of the same underlying world state [1][2]. Odyssey demonstrated it with a deathmatch built on GoldenEye 007, the 1997 Nintendo 64 first-person shooter, and published the demonstration as a browser-based preview at agora.odyssey.ml [1].

Odyssey calls Agora-1 "the first in a series of multi-agent world models" and "an early research preview", not a product [1]. The technical argument behind it is architectural. Most video world models fuse simulation and rendering into a single network that predicts the next frame from previous frames plus user input. Agora-1 splits those two jobs apart: one model learns how the shared world state evolves in response to player actions, and a second, [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) based model learns to render that state into pixels from any viewpoint [1][11].

## Release and access

| Item | Detail |
| --- | --- |
| Announced | 18 May 2026, in a post bylined by Odyssey co-founder and CEO Oliver Cameron [1][10] |
| Framing | "Early research preview", first in a planned series of multi-agent world models [1] |
| Participants | Up to four per shared simulation, human or AI [1] |
| Demo environment | Deathmatch built on GoldenEye 007 [1] |
| Preview URL | agora.odyssey.ml [1][15] |
| Companion releases | PROWL-1, announced 12 May 2026 [7]; Starchild-1, announced the same week [12] |

Odyssey's research index lists Agora-1 with a "Try Agora-1" link to the playable preview, while its neighbours PROWL-1 and [Starchild-1](https://aiwiki.ai/wiki/starchild_1) are listed with links to a paper and a technical report respectively [16]. Tech Times noted at launch that Agora-1's claims "come from Odyssey's own blog post and have not been independently benchmarked", with no technical paper or peer review attached [11]. As of 16 September 2026 the agora.odyssey.ml address still resolved and served a page titled "Agora-1".

## Architecture

Agora-1 learns two distinct functions instead of one [1].

The first is a state model. Odyssey trained it directly on the internal state of the source game, so that it learns gameplay dynamics and how state transitions follow from player actions [1]. Odyssey's own illustration of the shared state describes it as tracking each agent's health, position "and more" [1]. Because this state is explicit and discrete rather than latent, it can be edited directly: Odyssey says the consequence is that Agora-1 can generate entirely new levels while preserving gameplay dynamics consistent with the source game [1].

The second is a renderer. This is a DiT-based world model conditioned directly on the shared game state rather than on prompts, images, or the other conditioning signals typical of video world models [1]. It has to produce visually consistent renderings of the same state from several independent viewpoints at once, one per participant [1].

Odyssey describes the split as "loosely analogous to the structure of a modern game engine", with the difference that both halves are learned from data rather than hand-authored, and calls the result a "learned game engine" [1]. That phrase is the company's own; Agora-1 is not a game engine in the conventional sense and no hand-coded gameplay logic or rasterising renderer is involved [1][11].

The company is candid that each half poses its own research problem. Discrete game state is structurally different from the continuous visual domains that DiT-based world models usually operate over, which Odyssey says requires architectures designed specifically for gameplay state modelling and large amounts of structured training data [1]. Odyssey also describes the current state model as "relatively simple" and argues this is not a fundamental limit, expecting the state representation to scale to more complex simulations and eventually to generalise across rules and state representations [1].

Compared with clip-generating video models such as [Sora 2](https://aiwiki.ai/wiki/sora_2) or [Veo 3](https://aiwiki.ai/wiki/veo_3), which produce a fixed-length output that cannot be steered once generation begins, Agora-1 continuously advances a simulation state and renders a separate perspective for each player [10][12].

## Relation to earlier multi-agent world models

Odyssey's announcement explicitly names three earlier approaches to multi-agent interaction in world models, which means Agora-1 is not presented as the first multi-agent world model [1]. All three are real, published systems. The characterisations below are Odyssey's; where the original work describes itself differently, that is noted.

| System | Origin | What it actually does |
| --- | --- | --- |
| Multiverse | Enigma, blog post, 8 May 2025 [3] | Two-player world model trained on Gran Turismo 4 (a 1-v-1 race on the Tsukuba circuit), merging both player views into one input for a single U-Net diffusion model; released on GitHub and Hugging Face [3] |
| Solaris | Savva et al., arXiv:2602.22208, 25 February 2026 [4] | Multiplayer video world model for Minecraft; adapts a pre-trained single-agent video DiT so that players are interleaved along the sequence dimension and exchange information through a shared self-attention block; trained on 12.64 million multiplayer frames [4] |
| MultiGen | Po et al., arXiv:2603.06679, 3 March 2026 [5] | Adds an explicit external memory, a persistent state independent of the model's context window, and decomposes generation into Memory, Observation and Dynamics modules, giving editable environments and real-time multiplayer rollouts [5] |

Odyssey writes that Multiverse "concatenates agent states into a single 'split-screen' representation, effectively treating multiple players as one world state" [1]. Enigma's own write-up is more specific and partly at odds with the "split-screen" label: the team considered stacking the two player frames vertically, "just like classic split-screen games", but chose instead to stack them along the channel axis, treating both frames as one image with twice the colour channels, because the early convolutional layers of their U-Net only see nearby pixels: vertically stacked frames would not be processed together until the middle layers, whereas channel-stacked frames are processed together at every layer [3]. The Solaris paper independently describes Multiverse's design as "channel-concatenation" and calls it, as of February 2026, the only video world model it knew of that could simulate multiple agents [4].

On Solaris, Odyssey writes that it "concatenates each participant along the sequence dimension of a single autoregressive diffusion transformer, producing a more robust shared simulation", but that the approach "does not scale linearly with the number of players due to the growth of the model context" [1]. The first half matches the paper, which states that its modified DiT block "achieves multiplayer modeling through visual interleaving along the sequence dimension" [4]. The scaling criticism is Odyssey's assessment, not a claim made in the Solaris paper.

On MultiGen, Odyssey says only that Agora-1 is "similar to MultiGen" in maintaining an explicit shared world state, while adopting "a different approach to modeling simulation dynamics and rendering from that shared state" [1]. A secondary account of the launch went further and said MultiGen "doesn't separate simulation from rendering" [12]; the MultiGen abstract does describe a decomposition into Memory, Observation and Dynamics modules, so that stronger paraphrase should be treated with caution [5].

Odyssey's stated differentiator is consistency when participants lose sight of one another. It writes that both Multiverse and Solaris "struggle to robustly maintain consistency when players lose sight of one another", and that decoupling simulation from rendering lets Agora-1 produce consistent views of one world from multiple independent viewpoints [1][10].

## GoldenEye 007 as the test environment

Odyssey chose GoldenEye because it is "a game many on the Odyssey team loved growing up", and placed it in a lineage of games used as AI research environments alongside Atari, Minecraft and StarCraft [1]. The post's own citations for that lineage are the Nature papers on planning with a learned model across Atari, Go, chess and shogi [13] and on reaching Grandmaster level in StarCraft II [14], plus Odyssey's own Minecraft work on PROWL-1 [7].

GoldenEye 007 was released for the Nintendo 64 in 1997, developed by Rare and published by Nintendo, and includes a two to four player split-screen multiplayer mode [9]. Nintendo's copyright notice for the game names Nintendo and Rare for the game itself and Danjaq, LLC and Metro-Goldwyn-Mayer Studios for the James Bond material [9]. Microsoft bought Rare in September 2002 for $375 million, a deal whose announcement named GoldenEye 007 among Rare's games [17], so the game's developer now sits inside Microsoft while Nintendo remains its publisher [9][17].

That ownership tangle is why the choice of test bed drew comment. Tech Times called it "the most journalistically notable choice" in the launch, observing that Odyssey trained Agora-1 directly on the game's internal state and that "Odyssey's announcement does not address licensing or whether it has rights to the GoldenEye material it trained on" [11]. The outlet placed this against a field already in dispute over generated game worlds, noting that testers of Google's Project Genie produced playable knockoffs of Nintendo titles before Google restricted Nintendo content [11]. Odyssey's post itself makes no statement about the licensing of the source game [1].

## Multi-agent reinforcement learning

Odyssey's stated motivation for Agora-1 is less about games than about generating training experience. The company argues that progress toward more general agents is "increasingly bottlenecked not by model architecture, but by the experiences available during training", and that a world model supporting only one participant limits the kinds of [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) environments it can host [1].

The argument runs as follows. As participants are added, the joint interaction space grows combinatorially, so passively collected demonstrations cover a shrinking fraction of the interactions that matter: collisions, coordinated movement, contested objectives and other emergent behaviour [1]. Multi-agent RL inside a generated world is Odyssey's proposed mechanism for manufacturing that missing data, with agents and world models co-evolving and "continuously pushing one another into increasingly difficult regimes" [1].

This connects Agora-1 directly to [PROWL-1](https://aiwiki.ai/wiki/prowl_1), Odyssey's adversarial framework released six days earlier, in which an RL agent is rewarded for finding failures in a world model and those failures become fine-tuning data [7]. The accompanying paper, PROWL: Prioritized Regret-Driven Optimization for World Model Learning, implements the method in the MineRL framework and introduces a Prioritized Adversarial Trajectory buffer that re-ranks trajectories by prediction error, action fidelity and learning progress [6]. PROWL-1 is a single-agent loop; Odyssey presents Agora-1 as the system that removes that restriction [1].

A longer essay published by Odyssey researcher Ahmet Hamdi Guzel on 12 June 2026 sets out the surrounding research programme [8]. It situates Agora-1 in the "era of experience" framing associated with David Silver and Richard Sutton, reviews learning inside learned simulators from the [Dreamer](https://aiwiki.ai/wiki/dreamer) line through IRIS and DIAMOND, and points at the multi-agent gap: the work extending world models to teams, such as MAMBA, MARIE, CoDreamer and MATWM, is, in the essay's words, "almost all" cooperative, "and almost none of it is built around competitive self-play" [8]. The essay also notes that a world model which assumes a single actor implicitly assumes the rest of the world holds still, an assumption that breaks as soon as other learning agents change their behaviour [8].

The essay's forward-looking proposal is a multi-agent version of unsupervised environment design, in which "the next generation of PROWL" searches for a team's strategic weaknesses rather than a model's visual ones, and the world model generates social challenges (difficult opponent teams, configurations engineered to break coordination) that a procedural level generator could not [8]. Odyssey describes a separate internal setup running multi-agent RL on the SMACv2 benchmark inside its own StarCraft II environment over PySC2 [8]. The precedents the essay cites for emergent coordination are the StarCraft II league agent [AlphaStar](https://aiwiki.ai/wiki/alphastar), the Dota 2 team [OpenAI Five](https://aiwiki.ai/wiki/openai_five) and DeepMind's Capture the Flag agents; it mentions [SIMA](https://aiwiki.ai/wiki/sima) separately as an example of an agent following language instructions across many commercial 3D games [8][14].

Odyssey also proposes Agora-1 as a generative multi-agent simulator in its own right: a learned cooperative and competitive environment where policies trained entirely inside generated worlds might generalise to unseen environments and unseen partners without access to the original game [1].

## Beyond games

Odyssey states that the architecture is not limited to games, citing collaborative robotics, where multiple robots must jointly reason about actions, space and interaction with one another, and multi-view simulation more generally [1]. The launch post lists gaming, robotics, defence, education and foundation models as the domains it expects multi-agent world models to touch [1].

Commentators read that emphasis as continuous with the founders' background. Odyssey's leadership comes from self-driving: Oliver Cameron previously founded Voyage, acquired by [Cruise](https://aiwiki.ai/wiki/cruise), and Jeff Hawke was a researcher at [Wayve](https://aiwiki.ai/wiki/wayve) [11]. Tech Times argued that the named ambitions, including agents trained entirely inside generated worlds that transfer to unseen environments, amount to "the self-driving-simulation playbook applied to a broader class of problems" [11], a use of world models closer to [sim-to-real transfer](https://aiwiki.ai/wiki/sim_to_real_transfer) and [physical AI](https://aiwiki.ai/wiki/physical_ai) than to entertainment.

## Reception

Coverage treated Agora-1 as an architectural result rather than a consumer release. The Decoder described it as "one of the first multi-agent world simulations" and noted that until then world models had mostly been limited to a single active player, contrasting Agora-1 with [Genie 3](https://aiwiki.ai/wiki/genie_3), which it called visually more impressive but single-user [10]. Air Street Press, whose parent Air Street Capital is listed among Odyssey's investors [20], summarised the three prior approaches Odyssey names and described Agora-1's closest analogue as "a modern game engine, only with both halves of the engine learned from data rather than hand-authored" [12].

Tech Times gave the most sceptical read, calling Agora-1 "a research demonstration with a sharp idea and unverified claims, dressed in 1997 nostalgia to make an abstract advance legible", while allowing that keeping a shared world coherent across separate viewpoints by separating state from rendering is "a real and non-trivial step" [11]. It placed Agora-1 in a crowded field including Genie 3, [World Labs](https://aiwiki.ai/wiki/world_labs), [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos), [V-JEPA 2](https://aiwiki.ai/wiki/v_jepa_2), Decart and Runway [11].

Odyssey's announcement post on X reported the release as a playable research preview simulating a multiplayer GoldenEye deathmatch [2]. Odyssey CTO Jeff Hawke presented the company's research, including this line of work, at RAAIS 2026 in London on 12 June 2026 [12][18].

Agora-1 remains part of Odyssey's published research line alongside [Odyssey-2](https://aiwiki.ai/wiki/odyssey_2), [Odyssey-3](https://aiwiki.ai/wiki/odyssey_3), Starchild-1, [CaliBench](https://aiwiki.ai/wiki/calibench) and PROWL-1 [16]. Odyssey's later essay "Our Path to Superintelligence" cites Agora-1 as the demonstration that multiple participants, human or AI, could share and interact within the same generated world [19].

## Team

Odyssey credits the following core contributors for Agora-1 [1]:

| Role | Names |
| --- | --- |
| Core contributors | Aravind Kaimal, James Grieve, Sirish Srinivasan, Vinh-Dieu Lam, Zygmunt Łenyk |
| Leadership | Jeff Hawke, Oliver Cameron |

A longer "full team" credit lists Ahmad Nazeri, Ahmet Hamdi Guzel, Amogh Adishesha, Andy Kolkhorst, Ben Graham, Derek Sarshad, Fabian Güra, Finley Code, Jenny Seidenschwarz, Jesse Allardice, Jessica Inman, Jonathan Sadeghi, Kaiwen Guo, Kristy McDonough, Nicolas Griffiths, Nima Rezaeian, Richard Shen, Robin Tweedie, Sarah King, Tobiah Rex and Vighnesh Birodkar [1]. Several of those names, including Ahmet H. Güzel, Jenny Seidenschwarz, Benjamin Graham and Jonathan Sadeghi, are also authors of the PROWL paper [6].

## References

1. Oliver Cameron, "Agora-1: The Multi-Agent World Model", Odyssey, 18 May 2026. https://odyssey.systems/introducing-agora-1
2. Odyssey (@odysseyml), "Introducing Agora-1, a multi-agent world model", X, 18 May 2026. https://x.com/odysseyml/status/2056427523668668643
3. Enigma, "Introducing Multiverse: the first AI Multiplayer World Model", 8 May 2025. https://www.enigma.inc/posts/introducing-multiverse
4. Georgy Savva, Oscar Michel, Daohan Lu, Suppakit Waiwitlikhit, Timothy Meehan, Dhairya Mishra, Srivats Poddar, Jack Lu, Saining Xie, "Solaris: Building a Multiplayer Video World Model in Minecraft", arXiv:2602.22208, 25 February 2026. https://arxiv.org/abs/2602.22208
5. Ryan Po, David Junhao Zhang, Amir Hertz, Gordon Wetzstein, Neal Wadhwa, Nataniel Ruiz, "MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines", arXiv:2603.06679, 3 March 2026. https://arxiv.org/abs/2603.06679
6. Ahmet H. Güzel, Jenny Seidenschwarz, Benjamin Graham, Jonathan Sadeghi, Jeffrey Hawke, Ilija Bogunovic, "PROWL: Prioritized Regret-Driven Optimization for World Model Learning", arXiv:2605.18803, 11 May 2026. https://arxiv.org/abs/2605.18803
7. Jeff Hawke, "Introducing PROWL-1: Learning Through Discovery", Odyssey, 12 May 2026. https://odyssey.systems/introducing-prowl-1
8. Ahmet Hamdi Guzel, "The Era of Multi-Agent Imagined Experience", Odyssey, 12 June 2026. https://odyssey.systems/the-era-of-multi-agent-imagined-experience
9. "GoldenEye 007 returns for Nintendo Switch Online + Expansion Pack members", Nintendo, 26 January 2023. https://www.nintendo.com/us/whatsnew/goldeneye-007-returns-for-nintendo-switch-online-expansion-pack-members/
10. Maximilian Schreiner, "Agora-1 turns the N64 classic GoldenEye into a playable AI simulation for four players", The Decoder, 19 May 2026. https://the-decoder.com/agora-1-turns-the-n64-classic-goldeneye-into-a-playable-ai-simulation-for-four-players/
11. Tom K., "Odyssey's Agora-1 Puts Four Players Inside the Same AI-Generated World, Built on a 1997 Shooter", Tech Times, 19 May 2026. https://www.techtimes.com/articles/316846/20260519/odysseys-agora-1-puts-four-players-inside-same-ai-generated-world-built-1997-shooter.htm
12. Air Street Press and Nathan Benaich, "From clip-makers to simulators: Odyssey's new world models", 19 May 2026. https://press.airstreet.com/p/odyssey-starchild-1-agora-1
13. "Mastering Atari, Go, chess and shogi by planning with a learned model", Nature, December 2020. https://www.nature.com/articles/s41586-020-03051-4
14. "Grandmaster level in StarCraft II using multi-agent reinforcement learning", Nature, November 2019. https://www.nature.com/articles/s41586-019-1724-z
15. Agora-1 research preview. https://agora.odyssey.ml
16. "A leading body of world model research", Odyssey research index. https://odyssey.systems/research
17. "Microsoft Acquires Video Game Powerhouse Rare Ltd.", Microsoft, 24 September 2002. https://news.microsoft.com/source/2002/09/24/microsoft-acquires-video-game-powerhouse-rare-ltd/
18. "A missing form of intelligence", Odyssey (Jeff Hawke at RAAIS 2026). https://odyssey.systems/raais-2026
19. "Our Path to Superintelligence", Odyssey. https://odyssey.systems/our-path-to-superintelligence
20. "Investment from NVIDIA and Samsung", Odyssey, 12 February 2026. https://odyssey.systems/investment-from-nvidia-and-samsung