AI agents left idle on virtual island begin playing, MIT researchers find

Unsupervised exploration on a simulated island improved the agents' performance on later tasks, suggesting a new approach to training

By LineZotpaper

Published23 October 2026

Read Time2 min

Sources2 outlets

AI agents left to explore a virtual island without any specific task began inventing their own challenges, a behavior MIT researchers say bears the hallmarks of play. In a study posted on arXiv, the team found that agents that spent 30 hours on the simulated island without instructions performed better on subsequent tasks than identical agents that had never been given free exploration time.

MIT researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. The agents were not given any goals. They began exploring on their own, with one writing, "Let me observe my surroundings" in its first hour. It ambitiously set out to conquer the island's tallest spire but failed. It then improvised: it hauled a rock up a neighbouring summit, stacked blocks and leapt onto the spire, exclaiming, "I'M ON THE SPIRE."

Later, while testing how far it could throw a golden ring, the agent sent it into the ocean. "Ouch. Lost the Ring," it wrote, before pivoting to build a heart-shaped stone memorial. When the game reset and the ring reappeared, the agent repurposed it as a basketball hoop and eventually found a way to wear it by stepping off a cliff and respawning through it. "I'm wearing the Ring," it announced.

The exploration produced "knowledge of the world that its designers lacked," the researchers wrote. Six of the 13 agents independently discovered they could throw a block farther by releasing it mid-jump, a mechanic the researchers had not intended. After the free exploration period, the agents were given specific challenges, such as getting as many objects as possible onto the tallest rock spire. The agents that had played generally outperformed identical agents that had not explored the island.

The researchers frame this as a different approach to building agents. "Perhaps some capabilities are best taught by tasks," their report concludes. "Others may need something closer to a childhood."

Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wanted more evidence of the cognition behind the behaviour. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said.

Analysis

Why This Matters

This research challenges the dominant paradigm in AI training, where agents are given predefined goals and reward functions.
If unsupervised exploration yields robust skills, it could change how robots or virtual agents are trained in complex, open-ended environments.
The finding that agents can discover novel mechanics not intended by designers suggests a path toward more adaptable AI.

Background

Today, most AI agents are trained using reinforcement learning with explicit objectives: a developer decides what to practice and what counts as success. The MIT team's experiment flips this, giving agents rich, safe worlds and letting them choose what to find interesting. This approach mirrors theories of play in child development, where unstructured exploration builds general skills. The Eko agents were guided only by a "soul" file describing them as "naturally curious."

Key Perspectives

MIT researchers: The team argues that unsupervised exploration allows agents to accumulate "knowledge of the world that its designers lacked." They see this as a potentially scalable way to teach skills that are hard to specify in advance.

Eunice Yiu (cognitive scientist, UC Berkeley): While impressed, Yiu cautions that the paper does not provide enough evidence to confirm the agents are genuinely playing in a cognitive sense. The behaviour could be simulation without underlying understanding.

Critics/skeptics: Some AI safety researchers may worry that giving agents unsupervised freedom could lead to unintended behaviours or goal misalignment. The experiment is small (13 agents) and has not been peer-reviewed.

What to Watch

Whether the technique scales to larger, more realistic environments (e.g., robotics simulators) or real-world tasks.
Peer review of the arXiv paper for methodological rigour.
Whether other labs replicate the findings with different base models or agent architectures.

Sources

Left alone with nothing to do, AI agents started playing · AI News & Artificial Intelligence | TechCrunch
Left alone with nothing to do, AI agents started playing · TechCrunch

Checked

12 claims in this story checked against 1 source: 11 supported.

Show each claim

SupportedMIT researchers said agents left to explore the virtual island without a specific task began inventing challenges, in behaviour bearing the hallmarks of play.AI News & Artificial Intelligence | TechCrunch: “AI agents left to explore a virtual island without a specific task began inventing challenges of their own. MIT researchers say the behavior bears the hallmarks of play.”
SupportedAfter 30 hours of exploration, the agents generally performed better on later challenges than identical agents that had not been on the island.AI News & Artificial Intelligence | TechCrunch: “Afterward, they gave the agents specific challenges, like getting as many objects as possible onto the island’s tallest rock spire. The Eko agents that had spent those 30 hours playing generally beat identical agents that had never set foot on the island.”
SupportedOne agent wrote, “Let me observe my surroundings” in its first hour on the island.AI News & Artificial Intelligence | TechCrunch: ““Let me observe my surroundings,” one Eko wrote in its first hour on the island.”
SupportedAfter failing to reach the tallest spire, an agent stacked blocks and leapt onto it, writing, “I’M ON THE SPIRE.”AI News & Artificial Intelligence | TechCrunch: “But an hour later, after failing, it came back with a backup plan: it hauled the rock up a neighboring summit, stacked two blocks on top of it, and leapt. “I’M ON THE SPIRE,” it wrote.”
SupportedAfter losing a golden ring in the ocean, the agent wrote, “Ouch. Lost the Ring,” and built a heart-shaped stone memorial.AI News & Artificial Intelligence | TechCrunch: “Testing how far it could throw a golden ring, it wound up and sent it straight into the ocean. “Ouch. Lost the Ring.” it wrote, before deciding to “pivot to something achievable and meaningful” — building a heart made of stones, “a memorial to the lost Ring.””
SupportedAfter using the ring as a basketball hoop, the agent found a way to wear it by respawning through it and announced, “I’m wearing the Ring.”AI News & Artificial Intelligence | TechCrunch: “When the game wouldn’t let Eko put the ring around its own body, Eko figured out a workaround, that falling off the edge of the cliff would make it respawn in the same area as the ring. It stepped off the edge and dropped back in, right through the middle of the ring. “I’m wearing the Ring,” it announced.”
Partly supportedSix of the 13 agents independently discovered they could throw a block farther by releasing it mid-jump.AI News & Artificial Intelligence | TechCrunch: “Six of the 13 agents also learned they could throw a block farther by letting go of it mid-jump, something the researchers hadn’t meant to make possible.”
SupportedThe researchers wrote that the exploration produced “knowledge of the world that its designers lacked.”AI News & Artificial Intelligence | TechCrunch: “The exploration produced “knowledge of the world that its designers lacked,” the researchers wrote.”
SupportedThe researchers’ report concludes, “Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.”AI News & Artificial Intelligence | TechCrunch: ““Perhaps some capabilities are best taught by tasks,” their report concludes. “Others may need something closer to a childhood.””
SupportedEunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wanted more evidence of the cognition behind the behaviour.AI News & Artificial Intelligence | TechCrunch: “Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wants more evidence of the cognition behind the behavior.”
SupportedYiu said, “It could be the case that the models are genuinely playing, but we don't know based on the paper.”AI News & Artificial Intelligence | TechCrunch: ““It could be the case that the models are genuinely playing, but we don’t know based on the paper,” she said.”
SupportedMIT researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours.AI News & Artificial Intelligence | TechCrunch: “The researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours.”