AI agents given no task learned through play-like behavior on a virtual island

What is changing?

An MIT preprint published on October 5 examines 13 Eko agents operating on the same virtual island for 30 hours without an explicit task, reward or activity list. The Claude Opus 4.7-based agents perceived the environment through text commands and had access to persistent memory.

The agents developed activities including climbing, stacking blocks into towers, drawing geometric patterns, improvised bowling and basketball, and simple physics experiments. The researchers evaluated these through behavioral signatures of play such as novelty, surprise, repetition and emergence; the study does not claim consciousness or enjoyment.

Across four measurable goals, agents with 30 hours of experience generally outperformed zero-hour counterparts. Memory interventions indicated that some skills transferred through written experience and that removing a false belief could improve performance.

Additional experiments with GPT-5.5 and Kimi K2.6-based agents found similar behaviors. However, the study has not been peer reviewed; samples are small, the environment is limited, and the findings do not directly transfer to real robots or general autonomy.