--- title: Odyssey's Agora-2 puts 20 agents in one shared world at once description: Odyssey released Agora-2, a multi-agent world model letting up to 20 humans and AI agents interact in a single shared simulated environment in real time.... canonical: https://daily.dev/posts/odyssey-s-agora-2-puts-20-agents-in-one-shared-world-at-once-3m6xbnplm twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Odyssey's Agora-2 puts 20 agents in one shared world at once | daily.dev og:description: Odyssey released Agora-2, a multi-agent world model letting up to 20 humans and AI agents interact in a single shared simulated environment in real time.... og:url: https://daily.dev/posts/odyssey-s-agora-2-puts-20-agents-in-one-shared-world-at-once-3m6xbnplm og:image: https://api.daily.dev/og/posts/3m6Xbnplm.png og:image:alt: Odyssey's Agora-2 puts 20 agents in one shared world at once og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Odyssey's Agora-2 puts 20 agents in one shared world at once **[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 0 comments ## Summary Odyssey released Agora-2, a multi-agent world model letting up to 20 humans and AI agents interact in a single shared simulated environment in real time. Unlike single-agent world models, it maintains explicit shared state tracking how each participant's actions affect all others, predicting how combined actions ripple through the environment using entity properties, recent actions, and geometry. Proposed use cases include training self-driving cars, robots, and cyber defense agents alongside adapting adversaries, plus studying agent collusion in a contained setting before such behavior appears in production. Community reaction has been enthusiastic, though open questions remain about scaling beyond 20 participants, geometry modeling in complex environments, and performance versus domain-specific simulators. ## Content Odyssey just shipped a research preview of Agora-2, a multi-agent world model that runs up to 20 humans and AI agents inside a single shared simulation in real time. Every frame is generated by the model, and every participant's actions change what everyone else sees. That's the setup. Here's why people are paying attention. The core technical move is an explicit shared state that tracks how participants affect each other, updated using entity properties, recent actions, and surrounding geometry. Agora-1 could handle a fraction of this. Agora-2 supports 5x more participants and spans multiple environments at once. But the reaction isn't really about the numbers. It's about what this unlocks for training. The argument making the rounds: most agent failures in robotics, self-driving, and cyber defense aren't caused by the environment. They're caused by other agents. You can't train for that in isolation. Agora-2 is, as one observer put it, "a gym for AI agents that happens to be playable" — some of the other players are agents with their own goals, reacting to every move you make. That framing is doing a lot of work. The interesting part isn't the game. It's that you can now study agent collusion, adversarial coordination, and emergent multi-agent behavior in simulation before any of it reaches production systems. Odyssey's stated endgame is a feedback loop: better agents improve the world model, and a better world model trains better agents. Whether that loop actually closes cleanly is the open question nobody's answered yet. The takes are cautiously bullish. "Watch this space closely" is the consensus vibe, which in AI discourse usually means "impressive demo, jury's still out on whether it scales to anything real." The cyber defense and robotics angles are getting the most serious engagement — those are domains where multi-agent dynamics genuinely matter and current simulators are weak. The playable research preview is live now if you want to poke at it yourself. ## Questions this post answers ### What is Agora-2 and how does it differ from single-agent world models? Agora-2 is a multi-agent world model from Odyssey that lets up to 20 humans and AI agents share one simulated environment in real time. Unlike single-agent models that run agents in parallel silos, it maintains an explicit shared state tracking how every participant affects every other one, predicting how combined actions ripple through the environment using entity properties, recent actions, and surrounding geometry. _Anyone building multi-agent training pipelines can follow architecture shifts like this one on daily.dev._ ### What are the practical use cases for multi-agent world models like Agora-2? Flagged use cases include training self-driving cars against other drivers, cyber defense agents that adapt against active adversaries, and robots that improve together in a shared environment rather than in isolation. Another use case is studying agent collusion in a contained space, letting researchers observe unintended coordination behavior before it appears in production systems. _Teams evaluating multi-agent training approaches track emerging tools like this via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 4 discussions and 20 comments across x (as of 2026-09-25). **TL;DR:** Reactions are curious and largely positive about the leap to 20 concurrent agents/humans in a persistent shared world, but many raise technical questions about latency, state synchronization, and cost at scale. **Sentiment:** 35% positive · 45% mixed · 20% skeptical **The case for** - Persistent world state (things stay where they were left) is seen as a genuine unlock over prior demos. - Framed as a cheap sandbox to observe emergent multi-agent behavior like collusion before it happens in production systems. - Seen as a useful testbed for real-world applications like retail simulation or coordination research. **The pushback** - State synchronization and conflict resolution across concurrent agents may bottleneck latency more than raw world-model quality. - Shipping low enough latency for humans not to notice gets expensive fast. - Unclear how the model handles the sim-to-real gap when deployed outside its training domain. - Inference cost per player-minute at 20 concurrent agents is questioned as unclear. **By community** - x (mixed): Curiosity and enthusiasm about persistent shared worlds is tempered by pointed technical questions about cost, latency, synchronization, and sim-to-real transfer. **Hottest debate:** Whether the real bottleneck is world-generation quality or state synchronization/latency across concurrently acting agents. **Open questions** - What is the inference cost per player-minute at 20 concurrent participants? - How is state synchronization and conflict resolution handled as more agents act concurrently? - How is the sim-to-real gap addressed when a world model trained on one domain is deployed in another? **Highlights** > @testingcatalog The interesting bottleneck is state synchronization, not world generation. Once 20 agents act concurrently, stale observations and conflict resolution will dominate latency before raw model quality does. > — [amir\_dor on x](https://x.com/amir_dor/status/2103177602173452451) > @testingcatalog Every frame from the model and 20 agents rewriting each other's reality is either the best multi-agent eval harness we've got or a very expensive way to discover emergent griefing. When you drop a coding agent into Agora-2 — what's the first failure mode you actually want to > — [dydx77 on x](https://x.com/dydx77/status/2103169927225970971) > @omarsar0 First use case that comes to mind: multi-agent market sims. Watch agents collude in a sandbox before you deploy them anywhere near real money. Cheap lessons in-sim or expensive lessons in production. Pick one. > — [robofinancebk on x](https://x.com/robofinancebk/status/2103188959526310334) > @trikcode multiplayer world models are cool until you try to ship latency low enough for humans not to notice. then it gets expensive fast > — [harleyfoote\_ on x](https://x.com/harleyfoote_/status/2103155607570108628) > @Ai_hebrew @testingcatalog We didn’t add items to this world (besides your weapon) but the world remains consistent over time. You can kill a monster or have another participant occupy a spot, and come back to find them in the same spot > — [alipanju\_ on x](https://x.com/alipanju_/status/2103220719945658402) **Source threads** - [x](https://x.com/testingcatalog/status/2103164517433606640) · 0 points · 10 comments - [x](https://x.com/rohanpaul_ai/status/2103159817548554505) · 0 points · 2 comments - [x](https://x.com/trikcode/status/2103152475494617169) · 0 points · 4 comments - [x](https://x.com/omarsar0/status/2103151581810000240) · 0 points · 4 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/odyssey-s-agora-2-puts-20-agents-in-one-shared-world-at-once-3m6xbnplm) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Odyssey's Agora-2 puts 20 agents in one shared world at once","url":"https://daily.dev/posts/odyssey-s-agora-2-puts-20-agents-in-one-shared-world-at-once-3m6xbnplm","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/odyssey-s-agora-2-puts-20-agents-in-one-shared-world-at-once-3m6xbnplm"},"datePublished":"2026-09-24T16:29:41.369Z","dateModified":"2026-09-25T01:36:44.433Z","description":"Odyssey released Agora-2, a multi-agent world model letting up to 20 humans and AI agents interact in a single shared simulated environment in real time....","image":"https://pbs.twimg.com/media/HS_qGrqbsAAdAGa.png","thumbnailUrl":"https://pbs.twimg.com/media/HS_qGrqbsAAdAGa.png","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/odyssey-s-agora-2-puts-20-agents-in-one-shared-world-at-once-3m6xbnplm","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,reinforcement-learning","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Odyssey's Agora-2 puts 20 agents in one shared world at once"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/odyssey-s-agora-2-puts-20-agents-in-one-shared-world-at-once-3m6xbnplm#faq","mainEntity":[{"@type":"Question","name":"What is Agora-2 and how does it differ from single-agent world models?","acceptedAnswer":{"@type":"Answer","text":"Agora-2 is a multi-agent world model from Odyssey that lets up to 20 humans and AI agents share one simulated environment in real time. Unlike single-agent models that run agents in parallel silos, it maintains an explicit shared state tracking how every participant affects every other one, predicting how combined actions ripple through the environment using entity properties, recent actions, and surrounding geometry. Anyone building multi-agent training pipelines can follow architecture shifts like this one on daily.dev."}},{"@type":"Question","name":"What are the practical use cases for multi-agent world models like Agora-2?","acceptedAnswer":{"@type":"Answer","text":"Flagged use cases include training self-driving cars against other drivers, cyber defense agents that adapt against active adversaries, and robots that improve together in a shared environment rather than in isolation. Another use case is studying agent collusion in a contained space, letting researchers observe unintended coordination behavior before it appears in production systems. Teams evaluating multi-agent training approaches track emerging tools like this via daily.dev."}}]} ```