Left alone with nothing to do, AI agents started playing
- ID
- 38545
- Status
- summarized
- Published
- 23 Oct 2026, 8:45 AM
- Fetched
- 23 Oct 2026, 9:46 AM
- Provider
- TechCrunch
- Category
- technology
- Original URL
- https://techcrunch.com/2026/10/22/left-alone-with-nothing-to-do-ai-agents-started-playing/
- Source URL
- https://techcrunch.com/feed/
Summary
- Score
- 5.5
- Created
- 23 Oct 2026, 9:46 AM
- Tags
- Audience
- ai_ml_learnersai_agent_usersdevelopers
What happened
MIT researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, on a simulated desert island for 30 hours with no assigned task, and the agents began inventing their own challenges — one tried to scale the tallest spire, failed, stacked rocks on a neighbouring summit and declared 'I'M ON THE SPIRE'; another threw a golden ring into the ocean and pivoted to building a heart out of stones. Afterwards, the agents that had spent those 30 hours playing generally outperformed identical agents that had never been on the island on set challenges such as getting as many objects as possible onto the tallest rock spire. The work is described in an arXiv preprint (Cloos, Norelli, Durbin, Andreas, Rus, Isola) that has not been peer reviewed.
Why it matters
The contrast is with how most teams build agents today: you pick the practice tasks and define what counts as success up front. This preprint suggests an alternative loop — give the agent a rich, bounded environment, let it choose what's interesting, and let knowledge accumulate from its own history. It is a 13-agent, 30-hour, non-peer-reviewed experiment with no published effect size, so don't rip up your eval harness; the practical move is to note the pattern and, if you run agent training or self-improvement loops, consider whether your sandbox is safe and rich enough to let the agent pick its own sub-goals. Nothing here changes a shipping decision this week.
Discussion angle
If you let an agent choose its own practice in a sandbox, what stops it from optimising for something you'd call a bug — and what would you actually measure to prove the 'play' time transferred to real tasks?