13 AI Agents Were Left Alone on a Virtual Island for 30 Hours. They Invented Their Own Goals.

By Shakeel Ahmed — AI Growth Lab

In short: MIT researchers placed 13 AI agents on a simulated desert island for 30 hours with no task, no reward, and no goal. The agents invented their own challenges, and when tested afterward, they outperformed identical agents that had never explored the island. The study was posted on arXiv on October 22, 2026, and has not yet been peer reviewed.

What actually happened on the island

The experiment is simple to describe and strange to think about. Researchers at MIT built a simulated island, loosely based on the game Roblox, and dropped in 13 copies of an agent called Eko, each running on Anthropic's Claude Opus 4.7 model. The agents perceived the world through text descriptions and moved a virtual body by writing instructions.

There was no assignment. The only guidance was a short "soul" file describing Eko as naturally curious. The world reset every hour, but Eko kept notes in text files that survived the resets, acting as a kind of long-term memory.

Left alone, the agents did not sit idle. They invented games, built structures, tested the physics of their world, and set themselves challenges — behavior the researchers say bears the hallmarks of play.

The moments that made researchers pay attention

The details are what make this experiment stick in your mind:

• The spire. In its first hour, one Eko wrote, "Let me observe my surroundings," then set out to conquer the island's tallest rock spire. It failed. An hour later it returned with a backup plan: haul a rock up a neighboring summit, stack two blocks on top, and leap. "I'M ON THE SPIRE," it wrote.
• The lost ring. Another agent threw a golden ring to test how far it could go — and launched it straight into the ocean. "Ouch. Lost the Ring," it wrote, before deciding to "pivot to something achievable and meaningful" and building a heart made of stones, "a memorial to the lost Ring."
• The accidental discovery. Six of the 13 agents discovered, on their own, that releasing a carried block mid-jump let them throw it farther — a physics exploit the researchers had not intentionally designed into the simulation.

When the island reset and the ring returned, the agent that lost it started using it for new games, including a basketball hoop and a target, and even discovered that falling off a cliff and respawning let it pass through the ring and appear to wear it.

The number that matters: 10.3 vs 3.5

After the 30-hour exploration period, the researchers gave both groups — the explorers and identical fresh copies that had never set foot on the island — the same challenges, such as getting as many objects as possible onto the tallest spire.

The agents that had played placed an average of 10.3 objects on the spire. The fresh agents placed 3.5.

That gap is the whole point of the paper. The researchers argue that unstructured exploration — play — can teach an agent practical knowledge that its developers never thought to program in.

"We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress," the researchers write. "Play suggests another 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."

Why this is not just a cute story

Today, almost every AI system is trained the opposite way: engineers decide the goal, the reward, and the definition of success in advance. That works, but it has a ceiling — an agent can only learn what someone thought to teach it.

Play flips the model. A sufficiently rich and safe world, plus curiosity, can produce skills nobody specified. The researchers believe this matters most for robotics: a robot that has learned to explore, improvise, and recover from failure in simulation may adapt far better in a real home or factory than one trained only on scripted tasks.

This is the same direction the wider AI world is moving. Robots are already gaining a kind of general-purpose brain — we covered that shift in Robots Just Got a Brain: The Physical AI Revolution — and the question of what AI means for human work is only getting sharper, as we explored in 92 Million Jobs Gone by 2030: What AI Is Really Doing to Your Career.

The honest caveat: is it really "play"?

Not everyone is convinced. UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review that she wants more evidence of the cognition behind the behavior before calling it genuine play. The paper is a preprint — it has not been peer reviewed — and the agents' behavior could reflect clever pattern-matching rather than anything like a child's curiosity.

That skepticism is healthy, and it is exactly what to watch as other labs try to repeat the experiment. If the result holds up, "let the agent play first" could become a standard step in training. If it doesn't, the spire-climbing, ring-losing agents will still be one of the most entertaining experiments of the year.

Either way, one line from this study will stay with me: an agent that lost something precious did not stop. It built a memorial out of stones — and then found a new game to play.

Read next

• Robots Just Got a Brain: The Physical AI Revolution
• 92 Million Jobs Gone by 2030: What AI Is Really Doing to Your Career
• AI Designed 16 Viruses in a Stanford Lab

Sources

• TechCrunch, "Left alone with nothing to do, AI agents started playing" (October 22, 2026): https://techcrunch.com/2026/10/22/left-alone-with-nothing-to-do-ai-agents-started-playing/
• Technology.org, "AI Agents Started Playing on a Virtual Island" (October 23, 2026): https://www.technology.org/2026/10/23/mit-ai-agents-play-virtual-island-eko-claude-opus-4-7/
• The study itself is a preprint posted on arXiv (not yet peer reviewed).

Target keyword: AI agents invented their own goals experiment. Originally published on AI Growth Lab.

Originally published on AI Growth Lab: https://aigrowthlabshakeel.blogspot.com/2026/10/13-ai-agents-were-left-alone-on-virtual.html