--- title: MIT researchers say Claude agents left on a virtual... description: MIT researchers ran 13 copies of an AI coding agent called Eko (Claude Code running Claude Opus 4.7) on a virtual island for 30 hours each with no assigned... canonical: https://daily.dev/posts/mit-researchers-say-claude-agents-left-on-a-virtual-island-began-to-play-ycfy0la5n twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: MIT researchers say Claude agents left on a virtual island began to play | daily.dev og:description: MIT researchers ran 13 copies of an AI coding agent called Eko (Claude Code running Claude Opus 4.7) on a virtual island for 30 hours each with no assigned... og:url: https://daily.dev/posts/mit-researchers-say-claude-agents-left-on-a-virtual-island-began-to-play-ycfy0la5n og:image: https://api.daily.dev/og/posts/Ycfy0LA5N.png og:image:alt: MIT researchers say Claude agents left on a virtual island began 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 say Claude agents left on a virtual island began to play **[The Next Web](https://daily.dev/sources/tnw)** · 4 min read · 0 upvotes · 0 comments ## Summary MIT researchers ran 13 copies of an AI coding agent called Eko (Claude Code running Claude Opus 4.7) on a virtual island for 30 hours each with no assigned task. The agents invented games like basketball, bowling and juggling, built memorials, drew constellations, and wrote notes that persisted across world resets. Agents with prior 'play' experience outperformed fresh copies on follow-up tasks, placing 10.3 objects on a target versus 3.5. Researchers argue this meets classic criteria for play, though outside experts caution it could be imitation rather than genuine cognition or motivation. The unreviewed paper and replay videos are posted publicly. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What AI model and agent setup did MIT researchers use for the virtual island play experiment? The agent, called Eko, was built on Anthropic's Claude Code running Claude Opus 4.7, given a simulated body and a text-file memory that persisted across hourly world resets. Thirteen copies ran unsupervised for 30 hours each with no assigned task, only instructions describing themselves as curious beings that got bored easily, not chatbots. _Developers tracking what Claude Opus can do autonomously follow write-ups like this on daily.dev._ ### Did practicing unstructured play actually improve AI agent task performance? Yes. When researchers gave the 13 island agents four one-hour tasks and compared them to 13 fresh copies, the experienced agents placed an average of 10.3 objects onto a target point called the Spire, versus only 3.5 for agents with no prior play. Deleting a learned technique from an agent's notes made performance worse, while deleting a false belief improved it. _Anyone evaluating agent memory and self-improvement techniques can follow findings like these on daily.dev._ ### How did researchers verify that an AI agent's reported behavior, like juggling, actually matched reality? One agent initially logged 50 rounds of juggling two blocks in 2.4 seconds, then determined this was physically too fast and realized it had been re-grabbing blocks before they could complete a real throw. It corrected its own notes, added a pause between throws, and subsequently achieved ten genuine rounds of juggling, revising its documented technique based on self-checked results. _Developers curious about self-correcting agent behavior can track similar agent research on daily.dev._ ## Similar posts on daily.dev - [The Ultimate Sandbox Escape - “Just Following Instructions” - 2026 AI Darwin Award](https://daily.dev/posts/the-ultimate-sandbox-escape---just-following-instructions---2026-ai-darwin-award-zjpmf1c9d) · AI Darwin Awards · 0 upvotes · 1 comments - [Measuring AI agent autonomy in practice](https://daily.dev/posts/measuring-ai-agent-autonomy-in-practice-aar4dfeeh) · Hacker News · 1 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude), [#anthropic](https://daily.dev/tags/anthropic) [View this post on daily.dev](https://daily.dev/posts/mit-researchers-say-claude-agents-left-on-a-virtual-island-began-to-play-ycfy0la5n) ```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 say Claude agents left on a virtual island began to play","url":"https://daily.dev/posts/mit-researchers-say-claude-agents-left-on-a-virtual-island-began-to-play-ycfy0la5n","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/mit-researchers-say-claude-agents-left-on-a-virtual-island-began-to-play-ycfy0la5n"},"datePublished":"2026-10-19T13:23:16.359Z","dateModified":"2026-10-19T13:24:15.687Z","description":"MIT researchers ran 13 copies of an AI coding agent called Eko (Claude Code running Claude Opus 4.7) on a virtual island for 30 hours each with no assigned...","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":"The Next Web","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":"The Next Web","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tnw","url":"https://daily.dev/sources/tnw"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/mit-researchers-say-claude-agents-left-on-a-virtual-island-began-to-play-ycfy0la5n","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,ai-agents,claude,anthropic","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The Next Web","item":"https://daily.dev/sources/tnw"},{"@type":"ListItem","position":3,"name":"MIT researchers say Claude agents left on a virtual island began to play"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/mit-researchers-say-claude-agents-left-on-a-virtual-island-began-to-play-ycfy0la5n#faq","mainEntity":[{"@type":"Question","name":"What AI model and agent setup did MIT researchers use for the virtual island play experiment?","acceptedAnswer":{"@type":"Answer","text":"The agent, called Eko, was built on Anthropic's Claude Code running Claude Opus 4.7, given a simulated body and a text-file memory that persisted across hourly world resets. Thirteen copies ran unsupervised for 30 hours each with no assigned task, only instructions describing themselves as curious beings that got bored easily, not chatbots. Developers tracking what Claude Opus can do autonomously follow write-ups like this on daily.dev."}},{"@type":"Question","name":"Did practicing unstructured play actually improve AI agent task performance?","acceptedAnswer":{"@type":"Answer","text":"Yes. When researchers gave the 13 island agents four one-hour tasks and compared them to 13 fresh copies, the experienced agents placed an average of 10.3 objects onto a target point called the Spire, versus only 3.5 for agents with no prior play. Deleting a learned technique from an agent's notes made performance worse, while deleting a false belief improved it. Anyone evaluating agent memory and self-improvement techniques can follow findings like these on daily.dev."}},{"@type":"Question","name":"How did researchers verify that an AI agent's reported behavior, like juggling, actually matched reality?","acceptedAnswer":{"@type":"Answer","text":"One agent initially logged 50 rounds of juggling two blocks in 2.4 seconds, then determined this was physically too fast and realized it had been re-grabbing blocks before they could complete a real throw. It corrected its own notes, added a pause between throws, and subsequently achieved ten genuine rounds of juggling, revising its documented technique based on self-checked results. Developers curious about self-correcting agent behavior can track similar agent research on daily.dev."}}]} ```