AI agents left alone on an island started playing
MIT researchers say unguided agents invented their own challenges, a behavior they link to play.
Curious#08
13 agents
AI agents left alone on a simulated island
Image: TechCrunch
13 agentsAI agents left alone on a simulated island
The short version
- The researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours.
- Afterward, the Eko agents that had spent those hours playing generally beat identical agents that had never set foot on the island at specific challenges.
- Six of the 13 agents learned they could throw a block farther by letting go of it mid-jump, something the researchers had not meant to make possible.
- The findings are described in a paper posted on arXiv that has not yet been peer reviewed.
Why it matters
The work suggests agents might learn useful skills by exploring a rich, safe world without a fixed goal, instead of practicing what developers choose in advance. The researchers see this as a possible route to skills that could one day control robots. Developers building agents should note that the exploration produced knowledge its designers lacked.
“Perhaps some capabilities are best taught by tasks”