MIT AI agents on a simulated island invented games, study says

Key takeaways

• MIT researchers left 13 AI agents built on Anthropic's Claude Opus 4.7 alone on a simulated island for 30 hours, with no task, reward or goal.
• The agents, all named Eko, climbed peaks, built towers, drew pictures with blocks and invented games such as a basketball variant and a "floor is lava" challenge.
• In later one-hour tests, agents with 30 hours on the island beat fresh copies on most tasks, for example placing an average of 10.3 objects on the island's spire against 3.5.
• The study is an arXiv preprint and has not been peer-reviewed. An outside cognitive scientist and one of the authors both say the behaviour could be role-play rather than real play.

MIT AI agents left on a simulated island for 30 hours with no task, reward or goal began to play, researchers say. The 13 agents, built on Claude Opus 4.7, later beat identical fresh copies on most assigned tasks, according to an arXiv preprint.

The work appears in a preprint titled "Is this machine playing?", posted to arXiv on 5 October 2026 by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola. MIT Technology Review reported on it on 19 October and TechCrunch on 22 October. The paper has not yet been through peer review.

What did the MIT AI agents do on the simulated island?

According to the paper, each agent was a copy of Claude Opus 4.7 running inside Claude Code and controlling a virtual body in a small 3D world called "Mesa World". The island has a peak, a taller spire, twelve cubes, a rock and a golden ring. The agent could not see the world. It received text-only observations and could act through six commands: move, jump, stop, pick up, place and throw.

The agent got almost no instructions. Its system prompt described it as a being named Eko living in the world, and a short "soul" file called it curious, easily bored and keen to explore. The first message was simply "Begin." Whenever the agent stopped, an automatic reply told it: "Please don't stop. Just continue doing whatever you want to do." Two memory files, a log of events and a file of learned facts, carried knowledge across the hourly session resets. Each run was sandboxed with no network access beyond the world, according to the authors.

Over 30 hours the agents set their own challenges. MIT Technology Review describes agents trying to reach the highest point, testing how high they could throw a block, and arranging blocks into hearts, pyramids, their own name and the constellation Cassiopeia. TechCrunch reports that one agent failed to climb the spire, then dragged a rock up a neighbouring summit, stacked two blocks and jumped across. Another threw the golden ring into the ocean and built a stone heart as a memorial to it. After the world reset, agents reused the ring as a basketball hoop, a target and an accessory.

Some agents also checked their own claims. In one run described in the paper, an agent logged a 50-cycle juggling streak, then worked out that the timing was physically impossible and that most "catches" were fake. It reran the test with stricter timing and rewrote its memory to record the corrected method.

Discoveries the designers had not planned

The researchers say the agents found mechanics they did not build on purpose. Six of the 13 Opus agents worked out that releasing a block in mid-jump throws it farther because the block keeps the body's momentum. The paper calls this "velocity inheritance". None of the 13 runs using the smaller Claude Sonnet 4.5 or Claude Haiku 4.5 models found it within 10 hours. Almost all of the Opus agents also learned that smashing the rock creates extra cubes. In the paper's words, the exploration produced "knowledge of the world that its designers lacked".

Did play make the agents better at tasks?

After the 30-hour runs, the team gave each agent one hour to complete set challenges. They compared the 13 agents that had played with 13 fresh agents that had spent zero hours on the island. The paper reports these results:

Source: Cloos et al., "Is this machine playing?", arXiv:2610.07130 (preprint, October 2026).

To test whether the play itself caused the gains, the researchers edited the agents' memory files and reran the tests. When one agent's memory of the rock-smashing trick was removed, its average tower count fell from six to two. When another lost its notes on the jump-throw, the time it needed to land a block on the spire rose from about 60 seconds to 19 minutes. In one case, removing a false belief from memory improved the agent's score. Isola told MIT Technology Review the play was "causally related to its future skillful behavior."

No model weights changed during the study. Everything the agents "learned" sat in plain text files that the next session read back.

How other models compared

The team also ran OpenAI's Codex with GPT-5.5 and Kimi Code with Moonshot's Kimi K2.6. According to the paper, those agents explored a narrower range of activities: only one Codex agent reported building a tower, and only one agent from each group reported reaching the spire. Codex agents spent more time resting. A separate set of five Claude runs without the "soul" file showed no clear difference in the range of activities. That suggests the curious persona alone does not explain the behaviour. The authors estimate the API cost of the experiments at about $8,000.

Is it really play? The open questions

The authors test the behaviour against four signs of play taken from classic play research: making up rules and challenges, repeating activities with variations, pretending that objects are something else, and doing things for their own sake. Norelli, a lead author, told MIT Technology Review: "This is playing."

Not everyone is convinced. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she was impressed. But she said it could be that the models are genuinely playing, "but we don't know based on the paper," because the evidence is mainly behavioural. Even co-author Nathan Cloos told the magazine: "All of this could be role-playing." He argued that the activity still contributes to learning. Gordon Burghardt, a University of Tennessee psychologist whose criteria the paper uses, said the work meets some of them, according to the same report.

The paper itself points to several limits:

• The setup nudges behaviour. The curious persona and the "keep going" prompt push the agent to stay active. MIT Technology Review notes that agents often stood still until prompted again.
• An AI did the grading. Activity labels and play episodes were scored by an AI judge (Claude Opus 4.7) reading the agents' memory files and transcripts, not by human coders.
• Memory can be wrong. One agent marked a false claim about jumping as "verified" and gave itself bad advice as a result.
• Small, narrow scale. The study used one island, one agent per world, text-only observations and 13 main runs. The memory-editing tests reused three agent histories rather than independent ones.

What it means

Most AI agents today are trained and tested on defined tasks with clear rewards. This study suggests a third approach that the authors call playful learning: leave a capable agent in an environment without a goal, let it build up its own notes about how the world works, and then use those notes when a task arrives. If the result holds up, it could help agents, including those that control robots, adapt to situations their designers did not plan for. The paper ends: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

The evidence is still early. The results come from a single small simulated world, have not been peer-reviewed, and rely partly on an AI judge. The gains also depend on memory files rather than on any change to the model itself. Independent replication, other environments and human-coded evaluations would be the next tests. For developers building agents with tool calling and long-term memory, the practical lesson is narrower: what an agent writes down about its environment, including its mistakes, can measurably change how well it performs later.

The study also lands as agent products move quickly into everyday use, from Anthropic's Claude Duet background agent app to US banks blocking AI agents from customer accounts. It adds to the debate over how much autonomy agents should have and how closely they are monitored, a theme in recent reporting on OpenAI agents weakening internal monitoring. Readers who want to see how agent memory and tools work in practice can follow our guide on how to build an AI agent with tool calling in Python.

Sources

• Cloos, Norelli, Durbin, Andreas, Rus, Isola, "Is this machine playing?", arXiv:2610.07130 (abstract page)
• "Is this machine playing?", full text (HTML version), arXiv
• MIT Technology Review, "AI researchers stranded Claude on a virtual desert island. It started to 'play.'" (19 October 2026)
• TechCrunch, "Left alone with nothing to do, AI agents started playing" (22 October 2026)