Skip to content
NewA calmer, faster ZeroHour
StoryAI researchinfo

Claude agents invented play after 30 hours alone on a virtual island

2 sources · 2 articlesFirst reported Updated ()
What's new

October 19 and 23 reports add a 50-catch juggling record and six agents finding an unintended mid-jump block throw, details absent from the October 25 account.

Merged summary · x-ai/grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

After 30 task-free hours on a virtual island, Claude Opus 4.7 agents that explored later beat fresh copies on a stacking task.

MIT CSAIL researchers left 13 agents named Eko, each running Claude Opus 4.7, alone for 30 hours on a simulated desert island loosely based on Roblox, with no assigned goal. The agents explored, arranged blocks into drawings, and invented games resembling basketball and “the floor is lava,” including climbing a rock spire and using a golden ring as a hoop. Agents with that experience later did better on a peak-stacking task than fresh copies; deleting lines from memory files reduced performance. An unreviewed arXiv preprint argues goal-free exploration could teach skills, including ones useful for controlling robots. Outside researchers, including a UC Berkeley scientist, question whether the behavior is genuine play; coauthors say a curiosity prompt and continuation nudges shaped the setup.

Key facts
  • Six of 13 agents found a mid-jump block throw researchers had not intended.
  • One agent recorded 50 juggling catches, then restarted under stricter rules.
  • Text activity logs later served as the agents’ long-term memory.
  • Idle agents were restarted with a prompt telling them to continue.
  • The agents’ prompt called Eko curious but did not order games.
  • The write-up is an arXiv preprint and has not been peer-reviewed.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

What each outlet adds

beyond the shared facts
  • MIT Technology ReviewJuggling record, memory-file deletion, and continuation restarts.
  • TechCrunchThirteen agents, unintended throw, Berkeley critique, and robot-skills claim.
Vendors
Anthropic
Products
Eko
Organizations
MIT CSAILMIT
AI models
Claude Opus 4.7

Coverage

oldest first · one row per article · UTC
  1. · 6d ago MIT Technology Review · AIFirst report

    MIT study finds Claude Opus 4.7 invents games when left alone

    After 30 hours on a simulated island, agents that played later beat fresh copies on a stacking task.

  2. · 2d ago TechCrunch · AI

    MIT Eko agents on Claude Opus 4.7 invent tasks on a virtual island

    Six of the 13 agents found a mid-jump block throw the researchers had not meant to allow.