Given no assigned task, AI agents showed play-like behavior in simulation

Self-directed activity, not an instruction-free experiment

An AI experiment described in a paper submitted to arXiv on October 5, 2026, reports that agents developed their own activities in a simulated environment without an assigned objective. The researchers describe behavior resembling play, rather than establishing that the systems experienced enjoyment.

The main experiment involved 13 independent instances of an agent architecture called Eko, each operating for 30 hours. According to the researchers, the agents explored terrain, constructed towers, created patterned arrangements and experimented with their surroundings. The report concerns behavior inside a simulation, not physical robots acting in the real world.

There is an important qualification to the absence of assigned tasks. The setup included a file presenting Eko as inquisitive and prone to boredom. An orchestration system also prompted the assistant to continue when it stopped. Those design choices left activities unspecified, but they were still forms of direction.

What the researchers mean by play

The authors interpret the activity through four behavioral features: inventing rules, repeating activities with changes, treating objects imaginatively and pursuing goals without remaining rigidly committed to them. Their claim is about observable patterns. Whether an agent feels amusement or has an inner experience of playing remains a separate, unanswered question.

The learning claim—and its limits

The researchers also report that experienced agents generally did better than inexperienced agents on later goal-directed evaluations. Their memory interventions link that advantage to stored information: deleting descriptions of useful techniques impaired relevant performance, while removing an acquired misconception improved it.

Analysis: The narrower finding is more useful than a claim that machines became bored and entertained themselves. The study describes a deliberately configured system generating activities and retaining information that could help or hinder later work. It does not establish that prompting was irrelevant, that all exploration improved performance, or that human-like feelings were involved. These are the authors’ reported findings, not results independently reproduced for this draft.