AI agents playing games is the question at the centre of a new experiment by MIT researchers. Thirteen AI agents, left alone on a simulated desert island with nothing to do, began inventing challenges of their own. The researchers say the behaviour bears the hallmarks of play, and the agents that explored later did better on set tasks than identical agents that never went to the island.
What the researchers did
The team placed 13 AI agents, each called Eko and running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. The work is described in a paper posted on arXiv, a website where researchers share studies before they go through peer review, meaning other scientists have not yet formally checked it. Afterwards, the agents were given specific goals, such as getting as many objects as possible onto the island’s tallest rock spire. The Eko agents that had spent their 30 hours exploring generally beat identical agents that had never set foot on the island.
The researchers contrast this with the usual approach, in which developers decide in advance what an AI should practise and how success is measured. They see open exploration without a goal as a different way for agents to build skills, particularly skills that could one day be used to control robots. In their view, the key is giving an agent a rich and safe world and letting it choose what is interesting.
Games, a lost ring and a workaround
In its first hour on the island, one Eko set out to climb the tallest spire and failed. It came back with a backup plan, hauling a rock up a neighbouring summit, stacking two blocks and leaping onto the spire, then announcing “I’M ON THE SPIRE.” Later, while testing how far it could throw a golden ring, it sent the ring into the ocean and wrote “Ouch. Lost the Ring.” It then built a heart made of stones as a memorial to the lost ring.
When the game reset and the ring reappeared, the agent used it as a basketball hoop, a target and eventually an accessory. The game would not let it put the ring around its own body, so it found a workaround: falling off the cliff returned it to the same area as the ring. It stepped off the edge, dropped back in and passed right through the ring. Six of the 13 agents also learned that letting go of a block mid-jump threw it farther, something the researchers had not meant to make possible.
The researchers wrote that the exploration produced “knowledge of the world that its designers lacked.” They also acknowledge that the setup nudged the agents towards humanlike behaviour. The master prompt for Eko, a “soul” file, described it as “naturally curious.”
Doubts and what comes next
Eunice Yiu, a cognitive scientist at UC Berkeley, said she was impressed but wants more evidence about 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,” she said. Whether AI agents playing games amounts to play in a human sense is therefore still an open question.
The researchers conclude that some capabilities may be best taught through tasks, while others “may need something closer to a childhood.” Because the study has not yet gone through peer review, its findings will need to be checked by other researchers and tested in different settings before they can be treated as settled.
Aasan alfaaz mein
MIT ke researchers ne 13 AI agents ko ek virtual island par 30 ghante akela chhor diya, aur unhon ne khud challenges aur khel banaye. Kuch agents jo is dauran explore karte rahe, woh baad ke kaam mein aam taur par behtar nikle. Abhi yeh early research hai, is liye aap ki roz marra ki zindagi par iska foran asar nahin hoga. Lekin future mein AI kaise seekhta hai, is par yeh ideas asar daal sakte hain.
