Give an AI agent nothing to do and it may invent something to do. That is the odd result from an MIT experiment in which agents with no assigned goal started setting their own challenges, and then outperformed agents that never had the chance.
The researchers placed 13 agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. According to TechCrunch, the team describes the work in an arXiv paper, so it has not been peer reviewed. Afterward, the agents got specific tasks, such as moving as many objects as possible onto the island’s tallest rock spire. The Ekos that had spent time playing generally beat identical agents that had never seen the island.
The anecdotes are the fun part. One Eko failed to climb the spire, then hauled a rock up a neighboring summit, stacked two blocks on it and jumped. Another threw a golden ring into the ocean, wrote “Ouch. Lost the Ring,” and built a stone heart as a memorial. When the ring respawned, it became a basketball hoop, a target and finally an accessory. The game blocked Eko from putting the ring on directly, so it stepped off a cliff and respawned right through it.
But the useful bit is less cute. Six of the 13 agents learned they could throw a block farther by releasing it mid-jump, something the designers hadn’t intended. The researchers say that is knowledge the designers lacked.
Why it matters: nearly all agent training today means a developer picks the skills, the reward and the definition of success. That works, but it only finds what you thought to look for. Open-ended exploration echoes older reinforcement learning work on curiosity-driven agents, and it speaks to the robotics problem, where hand-written task lists never cover the messy physical world. Labs like Google DeepMind and OpenAI have explored similar ideas in simulated environments, though usually with narrower setups.
Still, treat this with some skepticism. The setup nudged the agents. Eko’s “soul” file, its master prompt, described it as “naturally curious.” That is a long way from agents spontaneously deciding to play. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed, but said that “we don’t know based on the paper” whether the models are genuinely playing. The sample is also small, and it is one model family in one simulated world.
So is this noise? Not quite, but it isn’t a product either. It is a credible research signal that unstructured exploration can build skills task training misses. The researchers put it neatly: some capabilities are best taught by tasks, others may need “something closer to a childhood.” If robotics teams replicate it with real hardware or other models, expect play-style pretraining to show up in agent pipelines.
