Claude agents improved at island tasks after free exploration, MIT researchers say

Time spent inventing games on a virtual island helped Claude-based AI agents tackle later challenges, MIT researchers say. In an experiment involving 13 agents nicknamed Eko, exploration produced discoveries they could reuse. The finding suggests a role for open-ended activity in building useful skills, without settling whether the agents were genuinely playing.

The MIT Computer Science and Artificial Intelligence Laboratory team gave the agents a bounded world with movable blocks, a large rock, a golden ring and elevated terrain. One summit began just beyond their reach. Copies of Claude Opus 4.7 controlled virtual avatars by writing instructions, rather than simply describing what they might do.

The 30-hour runs were not one continuous software session. The authors’ supplementary dataset explains that the world and agent process restarted between hours, while each agent’s workspace carried forward. That workspace included files for recording events and durable knowledge, plus a file describing the agent’s identity. The setup let written discoveries survive a restart.

Eko was not explicitly told to invent games or explore the island. But it was not starting from a blank slate: an identity document described it as naturally curious and eager to explore. That distinction becomes important when interpreting the activities that followed.

The agents devised versions of basketball and “the floor is lava,” jumping between small blocks without touching the ground. Another arranged blocks into drawings, progressing from hearts and pyramids to its own name and constellations. These were activities the agents selected, not a list of researcher-assigned challenges.

One agent recorded 50 consecutive catches while juggling, then rejected its own record because the catches happened too quickly. It started again under stricter rules. For coauthor Antonio Norelli, that voluntary escalation was a telling feature: the agent set a challenge, achieved it and then made the activity harder.

Exploration also exposed a useful feature of the simulated physics. Six agents independently discovered that releasing a block while jumping sent it farther than an ordinary throw, according to Gizmodo. The researchers said this behavior had not been deliberately built in as a discovery for the agents to make.

The strongest evidence for useful learning came after exploration. Researchers gave the agents specific tasks, including getting as many objects as possible onto the island’s highest peak. Agents with 30 hours of island experience performed better than fresh copies without that experience, the team told MIT Technology Review.

Their notes mattered. The agents had logged activities and discoveries in text files serving as long-term memory. Deleting lines from those files degraded later performance. That intervention connected the retained record to subsequent success, going beyond the observation that the agents produced entertaining behavior during their time on the island.

The authors also provide materials for examining the runs beyond selected anecdotes. Their repository lists world-state recordings, agent-session logs and end-of-hour workspaces. The recordings contain enough information to reconstruct and replay the simulated world, while the session files capture messages, tool calls and tool outputs.

The paper has not been peer-reviewed. Its authors identify four features associated with animal and human play: self-invented rules, repetition with variations, treating objects as something else and pursuing goals for their own sake. Psychologist Gordon Burghardt agreed that the work met some of his criteria for play in living creatures.

UC Berkeley cognitive scientist Eunice Yiu was more cautious. She said behavior alone could not establish the underlying cognition or motivation: the agents might be imitating humans or following their prompts. Eko also sometimes stopped doing anything until researchers supplied a continuation prompt asking it to keep doing whatever it wanted.

Coauthor Nathan Cloos acknowledged that the behavior could be role-playing, while arguing that it still contributed to learning. The team’s proposal is to give agents rich, safe environments where exploration can accumulate useful knowledge. Whether that requires the word “play” remains disputed; the later task results are the more concrete finding.