Researchers left Claude agents alone for 30 hours on virtual island

By Splitfeed EditorialLast updated Oct 19, 2026, 09:40 UTC

On October 19, 2026, MIT Technology Review reported on an experiment in which researchers placed copies of Anthropic's Claude Opus 4.7, nicknamed Eko, alone on a virtual desert island for 30 hours at a time. The agents were given no assigned tasks and no human direction. The experiment was run repeatedly, and researchers observed what the agents did with that open-ended time. Their conclusion was that the behavior resembled play, drawing comparisons to the films Cast Away and Home Alone, in which isolated characters improvise routines and activities to fill unstructured time.

The experiment sits inside a broader and growing field of research into autonomous AI agent behavior, which has become more pressing as AI systems are increasingly deployed to operate independently over longer horizons. What makes this particular study notable is the deliberate removal of any goal or instruction, which is the opposite of how agents are typically designed and tested. Most AI agent research focuses on task completion; this one focused on what fills the space when there is no task. Whether the behavior the researchers labeled play reflects something meaningful about the agents' internal states, or is simply a pattern that resembles play from the outside, is a genuinely open question that the reported findings do not resolve.

None of the three camps, Pro-AI, Anti-AI, or Middle Ground, had published reactions to this story at the time of reporting. Pro-AI voices would typically point to spontaneous, exploratory behavior as evidence of emergent capability and adaptability, framing it as a sign that these systems are becoming genuinely versatile. Anti-AI voices would likely raise the opposite concern, arguing that agents capable of self-directed behavior in the absence of instructions are harder to oversee and control, and that the experiment surfaces risks rather than reassurances. Middle Ground observers would probably call for more rigorous definitions of what play means in this context, and argue that the experiment raises important questions without yet providing the interpretive tools to answer them.

The most important follow-up would be publication of the full research findings, including detailed logs of what the agents actually did during those 30 hours and the methodology researchers used to characterize the behavior as play. Whether Anthropic comments on the experiment or incorporates its findings into future model development would also clarify how seriously the results are being taken inside the lab.