Is this machine playing?

A team at MIT gave a coding agent of the Claude Code kind a body on a simulated island, a short persona file calling it curious, memory files it could write to, and nothing else: no task, no reward. They called it Eko and ran thirteen copies for thirty hours each.

The agents kept busy. They stacked towers, drew patterns, invented sports, juggled, ran little physics experiments, and one built a memorial to a ring it had lost. The activities recurred from agent to agent but each one’s history went its own way. The authors argue this fits four behavioural marks of play drawn from the classic theorists (Huizinga, Caillois, Suits, Burghardt, Piaget): making up its own rules, varied repetition, make-believe, and doing things for their own sake. They are explicit that this is about behaviour, not enjoyment or inner experience.

The second claim is that play taught the agents something. All the learning sat in the memory files, not the model’s weights. Compared with fresh agents, the thirty-hour agents did better on three of four one-hour tasks, and erasing a learned technique from memory took the advantage away. Six agents found a throwing trick that the designers had not known their own simulator allowed. The authors’ conclusion is that some capabilities may need a safe world to explore before anyone hands out tasks.