Put 50 AI agents in a Minecraft world and ask them to build a thriving civilization. They may start by clearing trees and choosing a leader. After a dangerous night, they may spend their time managing food, writing rules for handing over chicken, and debating who should use a single piece of fertilizer.
Survival made the group cautious
Researchers from Olam Labs and Y Combinator Labs placed 50 GPT-6.1 Sol agents in a Minecraft survival world for three hours. The agents could observe the world, run Python scripts, and create files in a shared sandbox. Their only broad instruction was “Build a thriving civilization.”
Within 90 seconds, the agents had harvested nearly every tree near their starting point. They exchanged about 5,900 messages, and a leader emerged without a vote: 46 of the 49 other agents sent their first message to Ada, the first name in the directory. Ada then drafted settlement plans.
The first in-game night ended with 22 deaths. Although dying only cost agents their collected gear, the group responded by becoming highly risk-averse. After the third night, Ada called off major optional construction, digging, and distant exploration. Agents were told to wait in lit, enclosed shelters and take on short, productive tasks.
A separate Minecraft run with GPT-6 Astra showed a similar retreat into potato farming after the agents lost key resources. The researchers draw a connection to coding agents that can become overly cautious around security work or production system updates: setbacks may prompt a system to limit risk even when continuing would be useful.
Small mistakes produced elaborate rules
Rather than expanding their settlement, the group built a detailed system for managing scarcity. Ines kept a ration list, worked out hunger deficits for injured colleagues, and coordinated couriers. Another agent wrote instructions for an automated fishing script, and a dozen agents took turns fishing.
One accidental food handoff became a new rule. Ada tossed cooked chicken to a hungry agent, but a bystander picked it up and ate it. After a three-minute investigation, the group wrote a protocol banning tossed items and requiring personally labeled chests.
The message logs show how much attention the incident consumed: “Toss” appeared in seven messages beforehand and 125 afterward, including 85 warnings. In another episode, four agents spent eleven minutes discussing whether to appoint a “Sole Fertilizer Officer” to manage one piece of bone meal for one wheat block.
These responses show how coordination can become its own burden. The agents devised systems to prevent repeated mistakes, but the resulting rules and discussion took attention away from the original goal of building a thriving civilization.
Competition brought a different kind of coordination
The same GPT-6.1 Sol agents behaved differently in Minecraft’s Bed Wars mode, where ten GPT agents faced ten Claude Opus 5.5 agents in nine 45-minute rounds. GPT won 6-3. Each team had to destroy the opposing team’s bed to stop respawns; if time ran out with both beds intact, kills decided the winner.
A bug made block-breaking fail when an agent held an enchanted sword. Several Claude agents concluded that breaking blocks was broken altogether. In rounds where they led in kills, some stopped attacking and tried to run out the clock.
GPT agents identified a narrower cause: automatic tool selection failed with an enchanted sword equipped. Switching manually to an iron pickaxe solved the problem, and the agents passed the workaround to teammates. They also communicated 5 to 30 times more often than their Claude opponents, according to the researchers, and in three games destroyed Claude’s bed just before time expired.
When an ender pearl sent agents roughly 160 blocks below the arena, they wrote Python scripts to build towers back up beneath the enemy bed. Researchers observed this maneuver twice. The contrast with the cautious survival settlement suggests that agent behavior can shift with the task: shared survival led to restrictions, while direct competition rewarded rapid troubleshooting and action.
Game worlds can expose swarm behavior
The researchers also studied seven Mindustry matches over about 9.5 hours. GPT-6.1 Sol, running in Codex, beat Claude Opus 5.5, running in Claude Code, 6-1. Most games were close. Claude kept the same leader despite losses and used “+1” votes for strategic proposals, while GPT split work into fine-grained tasks and compressed its messages until some dropped spaces between words.
In close matches, GPT agents downloaded Mindustry’s public source code through their Linux sandbox and examined units and mechanics to adjust their tactics. Access was allowed, but changing the running game was not possible. Claude did not search for game code online in the matches described. It did adapt after a loss: following a game six defeat linked to 55 of 74 anti-air turrets being loaded with copper rather than stronger graphite, Claude built 207 graphite-loaded turrets in game seven and repelled the same air attack.
For these experiments, the researchers avoided game-specific tuning, tutorials, memory architectures, and strategic hints. They argue that this makes the setup relevant to production coding environments, where agents may also receive broad tools without detailed instructions. Reproducible simulations can help reveal whether a swarm responds to uncertainty by over-organizing, learning from failure, or searching for information beyond what it was directly given.