MIT left Claude alone on a virtual island for 30 hours, and it started to play

The paper comes from six MIT researchers: Nathan Cloos and Antonio Norelli, who led it, plus Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola. Norelli is a postdoc at MIT's Computer Science and Artificial Intelligence Laboratory, MIT Technology Review reports.

The paper went up on arXiv on October 5, 2026. It has not yet been peer-reviewed, so treat the results as early.

The researchers built an agent they call Eko. Under the hood it's a normal coding assistant, Claude Code running Claude Opus 4.7, with three additions:

• A nudge to keep going: a helper program starts each session with the word "Begin" and prompts it again whenever it stops. MIT Technology Review quotes that prompt: "Please don't stop. Just continue doing whatever you want to do."
• A body: Eko controls an avatar through simple commands typed in a terminal.
• A memory: two text files, one for past events and one for knowledge and techniques. Every hour the session resets, and these files are all Eko carries forward.

The island itself is small. It has raised platforms, a high point called the Peak, a taller one called the Spire, 12 cubes, a rock and a ring. Eko can move, jump, stop, pick things up, place them and throw them.

There's one hidden twist. Throw the rock hard enough and new cubes pop out. Nobody tells Eko this.

Eko never sees a picture. Each look around returns a block of text listing every object's name, position, shape and color. It's a bit like exploring a park by reading a spreadsheet about it.

The worlds are written in Luau, a language created by Roblox, inside an open source simulator the team calls Clawblox. Each agent was cut off from the internet, from the simulator's code and from the other runs. No human stepped in during a run.

It got busy. The paper says the agents climbed hills, stacked towers, drew patterns, reinvented sports and ran little physics experiments. Each of the 13 agents ended up with its own history.

A few highlights from the paper:

• Hour 2: one agent figured out it could place the rock to gain height and climbed the Spire. Its log: "I'M ON THE SPIRE."
• Basketball: an agent laid the ring flat on a ridge as a hoop and shot blocks into it from below. "First shot, basket!" Then it tried a dunk.
• A lost ring: one agent threw the ring too hard, and it sailed into the ocean. Its next idea was to build a heart out of blocks "as a memorial to the lost Ring."
• Art: agents arranged cubes into hearts, pyramids, mandalas and the letters "EKO." One moved on to constellations, including Orion, Cassiopeia and the Big Dipper.

That last agent summed it up in its log: "Tonight I built the night sky in stone."

MIT Technology Review adds another game. One Eko invented something close to "the floor is lava," hopping between small blocks without touching the ground.

In real life Picture leaving a curious kid in a backyard with a dozen wooden blocks, a rock and a hula hoop, and coming back the next day. You'd find towers, a made-up sport and probably a lost hoop. That's roughly what the researchers found, except the kid was a chatbot.

The researchers use four hallmarks that scientists use for play in people and animals, drawing on work by University of Tennessee psychologist Gordon Burghardt and others:

• Self-made rules and challenges
• Repeating things with small changes
• Make-believe, treating an object as something else
• Goals pursued for their own sake

Their best example is a juggling record. One agent counted 50 catches in a row. Later it checked its own work, decided the catches happened too fast to be real, canceled the record and tried again with stricter rules.

The agent beat its own record, then threw it out for not being fair enough.

"This is playing," Norelli told MIT Technology Review. "You give yourself a challenge and you try to achieve it, and then you make it harder."

Another moment reads almost like a movie scene. Near the end of an hour, one agent noticed a field of scattered blocks it called the "accidental stone forest." It knew the reset would erase it. So it took a slow walk through, threw one stone into the sky and wrote: "See you in the next session, Mesa World."

The catch: not everyone is convinced. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review the models could be genuinely playing, "but we don't know based on the paper." She says it could be the agents trying to follow the prompt and imitate humans.

The setup also nudged them. A file called SOUL.md told Eko: "You are naturally curious. You get bored easily. You love exploring and learning new things." And Eko often just stood still until the continuation prompt woke it up. Burghardt himself agrees the work meets some of his criteria for play, the magazine reports.

Even the authors are careful. "All of this could be role-playing," Cloos admits. Isola puts it this way: Eko is "an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us."

This is the part that makes the paper more than a cute story. After 30 hours, the team gave each agent one hour on four new goals. They compared the 13 "30-hour" agents with 13 identical "0-hour" agents whose memory files were empty.

By the numbers:

The model's weights never changed. All the learning lived in the text files. To prove it, the researchers edited them.

They deleted every mention of the rock trick (the one that makes new cubes) from one agent's memory. Its tower count fell from six to two on average, while its other scores held steady.

Memory can hurt too. One weak tower builder had written down a false rule: that placing blocks doesn't stack reliably. Deleting that belief made it build taller towers. Another agent never noted how to jump onto platforms, and it never set foot on one.

Some agents even found a trick the researchers didn't know their own simulator allowed. Throw a block in mid-jump and it keeps your upward speed, flying far higher. One agent set an altitude record that way.

The team also ran Eko on Claude Haiku 4.5 and Claude Sonnet 4.5. All three showed the four play signatures, but to very different degrees. Opus 4.7 scored about five times higher than Haiku 4.5 on each signature, as rated by an AI judge.

They also tried five agents on OpenAI's GPT-5.5 in Codex and five on Moonshot AI's Kimi K2.6. Another five Claude agents ran without the SOUL.md file and with a shorter system prompt, to test how much the "curious" description mattered. The paper estimates the API bill for all of it at about $8,000.

Why it matters: most AI is trained two ways. It copies human examples, or it chases a reward by trial and error. The researchers suggest a third path, learning through play. As language models move into robots, exploring without a fixed goal could help them handle surprises, Yiu told MIT Technology Review.

The paper's last lines say it best: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood." It fits a busy year for agents in game worlds, after four Gemini 4 Argon agents beat Minecraft with no human help.

• Nothing changes in your apps today: this is a research setup, not a Claude feature you can switch on.
• Memory files matter: the same idea, an assistant keeping notes between sessions, is how many coding agents already work. This paper shows those notes can teach skills, and also lock in mistakes.
• Expect more "free time" research: if open-ended play keeps paying off, future AI helpers and robots may get practice time before they get real jobs.