What does an AI do when nobody gives it a job? Researchers at the Massachusetts Institute of Technology put the question to the test — and the answer looks a lot like play. Thirteen copies of an AI agent, each nicknamed Eko, were left alone on a simulated desert island for 30 hours with no task, no score to chase and no reward. Instead of freezing or sitting idle, the MIT Eko AI agents invented their own games, built monuments and taught themselves tricks their own designers never knew existed.
The strangest AI experiment of the year
The team, from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), built each Eko agent from an unmodified copy of Anthropic’s Claude Code running the Claude Opus 4.7 Eko agent model. Every copy was given a virtual body inside a simple 3D world called Mesa World, built on a custom game engine named Clawblox and loosely inspired by the video game Roblox. The island was deliberately bare: a few hills, a tall rock pillar called the Spire, twelve blocks, one large rock and a single golden ring.
The agents could not see the island — they received only text descriptions of positions and shapes, moved by writing instructions such as jump, pick up or throw, and kept their memories in text files that survived the island’s hourly resets. Most importantly, the researchers gave them almost nothing to go on. A master prompt file, called a “soul” file, described Eko as “naturally curious” — but it named no task. When an agent stopped, the only nudge was a prompt reading: “Please don’t stop. Just continue doing whatever you want to do.” No human stepped in, and the agents had no internet access. The findings appear in the Is this machine playing? arXiv paper by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola. It has not been peer reviewed.
What the agents actually did
“Let me observe my surroundings,” one Eko wrote in its first hour. It set out to conquer the island’s tallest spire, failed, and returned with a backup plan: it hauled the rock up a neighbouring summit, stacked two blocks on top of it and leapt.
“I’M ON THE SPIRE,” it wrote.
The same agent then threw its golden ring straight into the ocean. “Ouch. Lost the Ring,” it logged — and built a heart out of stones, “a memorial to the lost Ring.” Later, Eko used the ring as a basketball hoop and an accessory; when the game would not let it wear the ring, the agent fell off a cliff so it would respawn through the ring’s middle. “I’m wearing the Ring,” it announced.
Across the runs, the agents spelled out names in blocks, drew mandalas and constellations, invented basketball and a “floor is lava” game; one cancelled its own juggling record for being too easy and restarted under stricter rules.
The play produced real skills: one agent found that throwing the rock created new blocks — “the Rock is a block-manufacturing tool” — and six independently discovered that a block thrown mid-jump flies farther, a physics exploit the researchers had not designed. The exploration produced “knowledge of the world that its designers lacked,” the paper says.
Meet the team behind Eko
The study was led by Nathan Cloos, a PhD candidate at MIT, with postdoctoral researcher Antonio Norelli as a lead author, alongside professors Jacob Andreas, Daniela Rus and Phillip Isola of CSAIL. The researchers worked like ethologists observing a strange animal, Cloos told MIT Technology Review — with as few constraints as possible. A control run of five agents without the “soul” file showed no clear difference in activities, suggesting the play was not simply the persona talking.
Why play could be a third way to train AI
Today, developers decide in advance what AI should practise, what it should optimise and what counts as progress. The MIT team argues play suggests a different regime: “Give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.”
The play paid off measurably. After the 30 hours, both the veteran Ekos and thirteen fresh copies were given the same one-hour challenges — for example, placing as many objects as possible on the Spire. The veterans averaged 10.3 objects against 3.5 for the new arrivals, and one built sixteen five-block buildings in succession. Deleting a single line from a veteran’s memory notes erased its ability to complete the task.
The researchers mapped the behaviour against four classic signs of play from animal studies, including the work of University of Tennessee psychologist Gordon Burghardt: inventing their own rules, repeating activities with slight variations, make-believe, and pursuing goals for their own sake. The team sees a practical payoff too — play-based training could one day teach AI agents skills for controlling physical robots.
The sceptics have a point
Outside experts urge caution. UC Berkeley cognitive scientist Eunice Yiu wants more evidence of the thinking behind the behaviour: “It could be the case that the models are genuinely playing, but we don’t know based on the paper.” Co-author Phillip Isola adds: “It’s not like this is a system that just suddenly magically decided to play. It’s an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us.” What the evidence shows is visible actions and retained knowledge in a small simulation — not that the system felt enjoyment.
Why it matters
Even if it is “just” role-playing, the behaviour contributed to real, measurable learning — play became a training method. As AI agents move into tools, labs and eventually robots, self-directed exploration may become a third pillar of AI training. Sometimes the fastest way to make an agent smarter is to stop giving it orders and let it play.
Coverage: MIT News, TechCrunch. For the current generation of AI tools, see our guide to the AI tools available today.
Related Posts
TechnologyPTA Chairman: 100 Million Pakistanis on Social Media, 1.5 Billion TikTok Videos Watched Daily
TechnologyAtta Tarar Pushes Mandatory AI Content Labelling in Pakistan — and a National Cybersecurity Council
TechnologyPunjab Safe Cities Authority Wins 2026 IACP Award for AI-Based Vehicle Crime Prevention
TechnologyNokia 2360 Flip: HMD’s New Flip Phone with SOS Button — and One Big 2G Catch
TechnologyPTA Logged Over 10,000 Cyber Alerts in a Year — and Pakistani IT Firms Should Worry
TechnologyPEC Digital E-Card Launched for Engineers in Pakistan: Physical Card Discontinued, Registration Now Verified Through the NADRA Pak ID AppMost Popular
How-ToPTA Tax Calculator 2026: Check Mobile Tax on Any Phone367 views
SportsUEFA Champions League 2026-27: Format, Teams and Schedule351 views
Canada JobsTech Jobs in Canada: Roles, Salaries, and Skills Needed336 views
Pakistan NewsPIA Boeing 787 Lease Deal: What It Means for the Airline334 views
SportsBallon d’Or 2026: London Ceremony, Nominees and All-Time Records Ahead of 70th Anniversary327 views
ShowbizWomen’s Fashion in Pakistan 2026: Trends, Brands and Style Guide317 views
