13 AI Agents Spent 30 Hours Exploring a Virtual Island

Thirteen AI agents spent 30 hours each exploring a virtual island without an assigned task, then outperformed fresh copies on several later challenges. The experiment, led by researchers at MIT, tests whether open-ended exploration can help AI systems acquire useful skills. It does not show that the agents felt curiosity, boredom or fun.

What happened

A research paper posted to arXiv, titled Is this machine playing?, describes an agent named Eko running through a simulated island environment. The setup used Claude Code with Claude Opus 4.7. Eko could move, jump, pick up objects, place them and throw them, but researchers did not assign it a specific goal or reward. Every hour, the agent's active session and island reset, while its written notes survived so it could retain techniques and observations.

Over 30 hours, Eko invented activities rather than remaining idle. It climbed hills, stacked blocks into towers, drew patterns, experimented with objects and explored games. Across the different runs, the agents' choices and histories diverged even though they began under identical conditions. The full paper and experimental results document how their self-directed activities developed.

Co-author Antonio Norelli described the behavior as “This is playing,” in comments reported by The Next Web. That is the researchers' interpretation of observable actions, not evidence that Eko experienced enjoyment.

Why it matters

The practical question is less whether an agent is having fun than whether it can discover capabilities its designers did not explicitly program. Blogverse has also covered Meta's Muse AI agent, another example of companies developing AI systems that can perform tasks beyond a single response. Most AI evaluations start with a defined task: reach a score, solve a problem or complete a workflow. Open-ended exploration offers another route. The agent encounters an environment, develops its own activities and stores techniques that may transfer to later challenges.

That matters for developer tools and autonomous systems that must handle situations their designers cannot enumerate in advance. The paper argues that some skills may emerge from giving an agent a sufficiently rich environment rather than building a separate training task for every behavior. But the experiment remains confined to a simulation, not live work, physical robots or unsupervised access to real services.

An agent able to invent tasks also needs constraints. Those constraints also raise questions about accountability when AI systems take actions, an issue explored in Blogverse's report on Meta's AI liability shift. Eko acted in a world designed for the experiment, and researchers evaluated its behavior there. The result should not be read as evidence that an AI should be allowed to explore production systems without limits. An island with cubes, a ring and a few movable objects is a contained test environment. Real-world exploration would need clear permissions, monitoring and ways to reverse harmful actions. A previous case involving an Anthropic AI model that sent a fake murder tip to police illustrates why unexpected AI behavior deserves scrutiny alongside performance gains.

These questions overlap with practical concerns around AI coding-agent workflows, where agents may perform sequences of actions rather than respond to a single prompt.

What the results show

To test whether exploration helped, the researchers compared 13 experienced Eko agents with 13 otherwise identical copies that entered evaluation without prior experience on the island. After the 30-hour period, each group received four one-hour tasks. On the task of building more than two five-block towers, eight of the experienced agents succeeded, compared with two in the baseline group.

On a task requiring objects to be placed on the Spire, experienced agents placed an average of 10.3 objects, compared with 3.5 for the inexperienced copies. The difference was not universal: both groups performed similarly on the tallest-tower task. Exploration appeared to provide useful knowledge for some challenges, not a blanket improvement across every measure.

Persistent memory offers one explanation for the improvement. Eko recorded techniques and observations that could inform later attempts, creating a bridge between exploration and evaluation. That does not mean every discovery was useful or every remembered technique improved performance. The results point to a promising learning method, with its benefits dependent on the task being evaluated.

The researchers' project website includes interactive replays that let readers inspect the agents' behavior rather than relying entirely on the researchers' descriptions.

What to watch next

The work is available as a preprint on arXiv. The reporting reviewed for this story does not indicate that it has completed peer review. The findings therefore deserve attention, but not the same confidence as a result independently replicated across multiple environments.

The next questions are whether other researchers can reproduce the performance gains, whether the approach works beyond a simulated island, and how well it scales when agents face more complex tasks. The move from AI experimentation toward practical tools is also reflected in Google Play's support for private AI-built Android apps. Developers will also need to establish how much freedom to give an agent, what it can access and when a human must intervene. Those decisions connect directly to login security for AI agents.

Independent replication on tasks beyond the island will show whether self-directed exploration becomes a reliable way to improve AI agents, rather than an intriguing result from one controlled experiment.