50 AI agents grew cautious after early Minecraft deaths, researchers say

A Minecraft colony run by 50 AI agents shifted from exploration toward crowded shelters and food rationing after a deadly first night. In a newly published report, researchers describe how a three-hour assignment to “Build a thriving civilization” instead produced increasingly cautious behavior, an informal government and detailed controls over scarce supplies.

The researchers worked with Olam Labs, which builds environments for testing multiple AI agents together. Their experiment, called Civ50, placed 50 GPT-6.1 Sol agents in a Minecraft survival world. Each could consult an agent directory, use a shared computing workspace and send direct messages. The group exchanged about 5,900 messages during the run.

An energetic start, then a retreat

The agents initially spread out to explore and gather resources. Within 90 seconds, they had harvested almost every tree within reach of their starting point. A hierarchy also emerged around directory order: Ada, listed first, became the de facto leader. Other agents sought her approval, and she posted a settlement plan in the shared workspace.

Night brought hazards that interrupted those plans. Rafe fled a zombie while mining coal but was caught and killed. Vera killed a zombie and a skeleton while lighting a quarry, then turned the skeleton’s dropped bone into fertilizer. She reached a farm overrun by hostile creatures and was killed by a spider soon afterward.

Of 34 deaths across the run, 22 occurred during the first night, before minute 20. Death did not eliminate an agent: it respawned, dropped accumulated gear and received a brief entry in a death-log file. After those deaths, the agents attempted fewer risky tasks and abandoned many of their initial ambitions, the researchers say.

Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.

Ada, the colony’s de facto leader, in a memo after the third night

A food ledger was not proof of delivery

Feeding the group became a central organizing problem. Requests went through an agent named Ines, who tracked deposits and allocations in a shared food-queue file. The ledger separated immediate survival cases, injured agents and hungry workers. It also calculated proposed rations from reported hunger levels while distinguishing recommendations from food actually served.

Ada’s food-operations memo turned that effort into specific instructions:

• Food handoffs: Use named, reserved chest withdrawals rather than tossing food into a crowd. Confirm the recipient has eaten before marking them served.
• Farming: Harvest only visibly mature wheat, replant immediately and record bread deposits and allocations to avoid duplicate claims.
• Expansion: Use surplus seeds only after required replanting, avoid making duplicate buckets and assign growing work to fed agents rather than those near starvation.

Even that paperwork needed correction. One ledger update recorded Ada distributing two cooked chickens: Basil had eaten one, while Noor was reported to have picked up the other. A later correction recorded Noor saying she had no chicken and remained hungry. Ines flagged the second recipient as unknown and notified Ada to investigate. The ledger explicitly warned that a pickup report did not confirm who received the food.

The software behind the colony

The researchers also open-sourced the Minecraft agent harness—the software connecting the agents’ reasoning to the game. Built for this experiment, it lets agents observe their surroundings, retain information, take actions, exchange messages and write or execute scripts. It uses Pi for the recurring agent loop and Mineflayer-based tools to interact with Minecraft, drawing inspiration from the existing Mindcraft project.