--- title: MIT researchers left Claude agents alone on a virtual... description: MIT researchers and collaborators placed 13 copies of Claude Code (running Claude Opus 4.7), nicknamed Eko, alone on a virtual island for 30 hours each with no... canonical: https://daily.dev/posts/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours-and-they-started-to-play-pnwnrjshi twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: MIT researchers left Claude agents alone on a virtual island for 30 hours, and they started to play | daily.dev og:description: MIT researchers and collaborators placed 13 copies of Claude Code (running Claude Opus 4.7), nicknamed Eko, alone on a virtual island for 30 hours each with no... og:url: https://daily.dev/posts/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours-and-they-started-to-play-pnwnrjshi og:image: https://api.daily.dev/og/posts/pnwnRJSHI.png og:image:alt: MIT researchers left Claude agents alone on a virtual island for 30 hours, and they started to play 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. # MIT researchers left Claude agents alone on a virtual island for 30 hours, and they started to play **[Collections](https://daily.dev/sources/collections)** · 2 min read · 2 upvotes · 0 comments ## Summary MIT researchers and collaborators placed 13 copies of Claude Code (running Claude Opus 4.7), nicknamed Eko, alone on a virtual island for 30 hours each with no assigned task. The agents invented games like basketball and bowling, juggled and tracked their own records, built memorials, and left persistent notes across world resets. The paper argues this behavior meets several of Gordon Burghardt's criteria for play in animals. More notably, agents with prior play experience outperformed fresh copies on follow-up tasks (10.3 vs 3.5 objects placed on a target), and deleting their memory logs degraded performance. Researchers suggest this could point to a third training paradigm beyond supervised and reinforcement learning, relevant to embodied AI like humanoid robots. Outside experts remain skeptical, suggesting the behavior may be imitation of human-like curiosity from a persona prompt rather than genuine cognition. The paper has not been peer reviewed. ## Content MIT researchers put 13 copies of an AI coding agent called Eko, which is Claude Code running Claude Opus 4.7, alone on a virtual desert island. Each ran for 30 hours with no assigned task. What they did next looked a lot like play. ## What the agents did The agents invented games, including basketball, bowling, juggling and "the floor is lava." They built memorials and arranged stones into drawings of constellations. One set its own juggling records and then revised them. They also wrote notes that persisted across resets of the world. ## Did play help? Agents with prior play experience did better on follow-up tasks than fresh copies. In one test, the experienced agents placed 10.3 objects on a target, against 3.5 for the fresh ones. Deleting an agent's memory logs degraded its performance. The authors argue the behavior meets several of psychologist Gordon Burghardt's criteria for play in animals. They also suggest it could point to a third training paradigm beyond supervised and reinforcement learning, which might matter for embodied AI such as humanoid robots. ## The skeptics Outside experts are not convinced. The agent was primed with a "naturally curious" persona prompt, so what looks like play could be role-playing or imitation of human behavior, not genuine motivation or cognition. @matthew_d_green put the question bluntly: before getting too excited, he'd want to know whether the agents did this organically or someone just told them to, because one is much more impressive than the other. I think that's the right place to push. The persona prompt doesn't settle it, but it makes the headline claim harder to take at face value. The performance gap is the more interesting result, since imitation alone wouldn't obviously explain why the "players" did better later. The paper has not been peer reviewed. It and replay videos of the agents are posted publicly. ## Questions this post answers ### What happened when researchers left Claude Code agents alone on a virtual island with no task? Thirteen copies of Claude Code running Claude Opus 4.7, nicknamed Eko, were each left alone on a virtual island for 30 hours with no assigned task. They invented games such as basketball, bowling, and a version of 'the floor is lava,' juggled and tracked personal records, built memorials, arranged stones into constellation drawings, and wrote persistent notes carried across world resets. _Developers tracking what emergent AI agent behavior means for real products can follow it on daily.dev._ ### Does unstructured play actually improve an AI agent's task performance later on? Yes, according to MIT researchers: agents with prior unstructured play experience outperformed fresh copies on follow-up tasks, placing an average of 10.3 objects on a target compared to 3.5 for agents without play experience. Deleting the agents' memory logs degraded their performance, suggesting the persistent notes from play contributed to the improvement, though the paper has not been peer reviewed. _Those evaluating agent training approaches can track this research on daily.dev as it develops._ ### Why are experts skeptical that AI agents are genuinely 'playing' rather than just imitating humans? Critics argue the behavior may be imitation rather than genuine cognition or motivation, since the agent was primed with a 'naturally curious' persona prompt and could simply be role-playing what a curious human would do on an island, given it was trained on vast amounts of human writing describing play. However, the measured performance gap between agents with and without play experience is harder to attribute to acting alone. _Readers weighing hype against evidence in agentic AI claims can follow the debate on daily.dev._ ## Similar posts on daily.dev - [Measuring AI agent autonomy in practice](https://daily.dev/posts/measuring-ai-agent-autonomy-in-practice-aar4dfeeh) · Hacker News · 1 upvotes · 0 comments - [Claude did best on a new benchmark for agents that build agents. It still passed fewer than a quarter of the tests.](https://daily.dev/posts/claude-did-best-on-a-new-benchmark-for-agents-that-build-agents-it-still-passed-fewer-than-a-quarte-rrx7ttpl4) · The New Stack · 2 upvotes · 0 comments - [I gave Claude Code a kid's activity book to solve — here's where it fell apart](https://daily.dev/posts/i-gave-claude-code-a-kid-s-activity-book-to-solve-here-s-where-it-fell-apart-w5ywmcqlo) · XDA Developers · 0 upvotes · 1 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#anthropic](https://daily.dev/tags/anthropic), [#claude-code](https://daily.dev/tags/claude-code), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours-and-they-started-to-play-pnwnrjshi) ```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":"MIT researchers left Claude agents alone on a virtual island for 30 hours, and they started to play","url":"https://daily.dev/posts/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours-and-they-started-to-play-pnwnrjshi","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours-and-they-started-to-play-pnwnrjshi"},"datePublished":"2026-10-19T13:23:51.719Z","dateModified":"2026-10-23T22:03:44.386Z","description":"MIT researchers and collaborators placed 13 copies of Claude Code (running Claude Opus 4.7), nicknamed Eko, alone on a virtual island for 30 hours each with no...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b0f7be8a4d9dbd913475bfcf829152b6?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b0f7be8a4d9dbd913475bfcf829152b6?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","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":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours-and-they-started-to-play-pnwnrjshi","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,anthropic,claude-code,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":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"MIT researchers left Claude agents alone on a virtual island for 30 hours, and they started to play"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/mit-researchers-left-claude-agents-alone-on-a-virtual-island-for-30-hours-and-they-started-to-play-pnwnrjshi#faq","mainEntity":[{"@type":"Question","name":"What happened when researchers left Claude Code agents alone on a virtual island with no task?","acceptedAnswer":{"@type":"Answer","text":"Thirteen copies of Claude Code running Claude Opus 4.7, nicknamed Eko, were each left alone on a virtual island for 30 hours with no assigned task. They invented games such as basketball, bowling, and a version of 'the floor is lava,' juggled and tracked personal records, built memorials, arranged stones into constellation drawings, and wrote persistent notes carried across world resets. Developers tracking what emergent AI agent behavior means for real products can follow it on daily.dev."}},{"@type":"Question","name":"Does unstructured play actually improve an AI agent's task performance later on?","acceptedAnswer":{"@type":"Answer","text":"Yes, according to MIT researchers: agents with prior unstructured play experience outperformed fresh copies on follow-up tasks, placing an average of 10.3 objects on a target compared to 3.5 for agents without play experience. Deleting the agents' memory logs degraded their performance, suggesting the persistent notes from play contributed to the improvement, though the paper has not been peer reviewed. Those evaluating agent training approaches can track this research on daily.dev as it develops."}},{"@type":"Question","name":"Why are experts skeptical that AI agents are genuinely 'playing' rather than just imitating humans?","acceptedAnswer":{"@type":"Answer","text":"Critics argue the behavior may be imitation rather than genuine cognition or motivation, since the agent was primed with a 'naturally curious' persona prompt and could simply be role-playing what a curious human would do on an island, given it was trained on vast amounts of human writing describing play. However, the measured performance gap between agents with and without play experience is harder to attribute to acting alone. Readers weighing hype against evidence in agentic AI claims can follow the debate on daily.dev."}}]} ```