
What happened
Olam Labs and Y Combinator put 50 GPT-6.1 Sol agents into Minecraft for 3 hours to build a thriving civilization; they hit food shortages, introduced rationing, and failed to meet the goal.
Why it matters
The agents organized themselves — an agent named Ada became de facto leader and issued rationing rules — yet still could not sustain the civilization, suggesting coordination alone did not solve scarce resources.
WHO IT HITSThis lands on teams evaluating AI agents for open-ended, multi-agent tasks — showing that even a flagship model can organize labor and rationing but still fall short on resource planning over a few hours.
Summaries like this, every morning.
The experiment is notable less for the failure itself than for how the agents got there. Early on, the 50 agents explored and felled nearly every tree within reach of the spawn point within 90 seconds, and an agent named Ada emerged as de facto leader without any discussion — 46 of the other 49 agents sent their first direct message to Ada. Ada posted a colony-building plan covering farmland, sapling protection, shelter coordinates and priority resources, and later issued rationing rules as food ran short after the third night.
The team also highlighted bureaucratic behavior: roles such as "fishing rod officer" appeared, low-level agents asked permission to harvest grass, and four agents spent 11 minutes applying a single bone meal to a single wheat field. Fishing became a key food source because it was low-risk, and a system of fishing, cooking, distribution and delivery emerged.
A second experiment pitted 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents in bed wars, with GPT-6.1 Sol winning 6-3. The same pattern of the first-listed agent becoming leader was observed there too, and the team says it is now running a 25-agent GPT-6.1 Terra experiment streamed on Twitch.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Microsoft detailed an Industry Evaluation Data Generation Platform that builds fully synthetic enterprise tena…

The Biological Computing Co

Anthropic's prospectus devotes 69,800 words, about 30% of its 240,000-word filing, to risks like job displacem…

Goldman economist Pierfrancesco Mei found AI-displaced workers since 2025 took about a month longer to re-land…

WTW's board approved "Propel," a company-wide AI and automation plan running through December 31, 2028, disclo…

Runware told TechCrunch it is building a Romania factory to assemble its Sonic Inference Pods on a production…
