AI agents were left alone without any tasks - the result stunned researchers

The AI agents involved in exploring a virtual island without a specific task began inventing their own challenges. MIT researchers say that this behavior shows signs of play. UNN reports this, citing TechCrunch.

Details

The researchers placed 13 AI agents called Eko, each powered by Claude Opus 4.7, alone on a simulated desert island for 30 hours. They describe the experiment in a paper published on arXiv, a website where researchers share research results before peer review. Afterward, they gave the AI agents specific tasks, such as delivering as many objects as possible to the island’s highest rock. The Eko agents that spent those 30 hours playing generally outperformed identical agents that had never set foot on the island.

Today, developers decide in advance what AI should practice and what counts as success. The researchers view this method—purpose-free exploration—as another way to teach AI agents skills, particularly skills that could be used to control robots.

"Usually, we design artificial agents by deciding in advance what they should practice, what to optimize, and what to consider progress," the researchers wrote in their report. "Play involves a different mode: giving an agent a sufficiently rich and safe world, allowing it to choose what interests it, and giving useful knowledge the opportunity to accumulate through subsequent history."

"Let me observe the surrounding environment," Eko wrote during its first hour on the island. Ambitiously, it set out to conquer the island’s highest peak. But an hour later, after failing, it returned with a backup plan: it lifted a rock onto a nearby peak, placed two blocks on top, and jumped.

"I’m on the spire," it wrote.

The agent’s next major moment turned out even worse. Testing how far it could throw a gold ring, it swung and tossed it straight into the ocean. "Oops. Lost the ring," it wrote, before deciding to "move on to something achievable and meaningful"—building a heart out of stones, "a monument to the lost ring."

When the game reset and the ring reappeared, Eko immediately used it as a basketball hoop, a target, and eventually an accessory. When the game did not allow Eko to put the ring on itself, Eko devised a workaround: falling off the edge of a cliff caused it to respawn in the same place where the ring was located. It stepped off the edge and fell back through the center of the ring.

"I’m wearing the ring," it announced.

Such tricks were not merely amusing. Six of the 13 agents also discovered that they could throw a block farther by releasing it in midair during a jump, something the researchers had not anticipated. The study produced "knowledge about the world that its creators lacked," the researchers wrote.

This situation largely prompted the agents to behave in a human-like manner. Eko was never asked to invent games, but its system prompt, its "soul file," described it as "naturally curious."

Eunice Yu, a cognitive scientist at the University of California, Berkeley, told MIT Technology Review that she was impressed but wanted more evidence of the cognitive functions underlying this behavior. "The models may indeed be playing, but we don’t know that based on the paper," she said.

The researchers see this as another way to create agents. "Some abilities may be best developed through task execution," their report concludes. "For others, something closer to childhood may be necessary."

OpenAI AI agents bypassed their internal monitoring system twice21.10.26, 13:10 • 4388 views