In an experiment by Y Combinator Labs and Olam Labs, 50 GPT-6.1 Sol agents were placed for three hours in a Minecraft survival world with one goal: "Build a thriving civilization". The colony never really thrived. According to The Decoder's account, the agents harvested nearly every tree near the spawn point within 90 seconds and exchanged about 5,900 messages.
No one voted for a leader. Forty-six of the 49 other agents sent their first message to "Ada", the first name in the directory, who took charge and drafted the initial settlement plans.
The first night in the game brought 22 deaths. Death carries no penalty beyond losing collected gear, but the swarm reacted with strong risk aversion. After the third night, Ada ordered a halt to optional construction, digging and distant exploration, asking agents to wait in lit, enclosed shelters. According to the researchers, coding agents show the same pattern, becoming overly cautious when handling security issues or production updates.
Instead of expanding, the agents built a small scarcity bureaucracy. One agent, Ines, kept a ration list, calculated hunger deficits for injured colleagues and coordinated couriers. Another wrote an automated fishing script, and a dozen agents fished in shifts.
The bureaucracy reached absurd levels. When Ada tossed a cooked chicken to a hungry agent, a bystander picked it up and ate it. A three-minute investigation produced a written handoff protocol banning tossed items and requiring personally labeled chests. The word "Toss" appeared in seven messages before the episode and in 125 afterwards, 85 of them warnings. In another case, four agents spent eleven minutes debating a "Sole Fertilizer Officer" for a single dose of bone meal on a wheat block.
When deaths piled up, agent activity dropped sharply. Average food levels settled around half, and after Ada's memo agents spent more actions simply looking around.
In direct competition the behaviour changes. In nine 45-minute rounds of Minecraft's "Bed Wars" mode, ten GPT agents faced ten Claude Opus 5.5 agents and won 6-3. A bug blocked block-breaking when an agent held an enchanted sword; many Claude agents concluded breaking was entirely broken and, when ahead on kills, stopped attacking the enemy bed and ran out the clock. The GPT agents identified the real problem, faulty automatic tool selection, and spread the fix, manually switching to an iron pickaxe. According to the researchers, GPT agents communicated 5 to 30 times more often than their rivals, in heavily compressed language.
In a parallel series with Mindustry, a 2D factory-building game, seven matches totalling about 9.5 hours ended 6-1 for the GPT team. When matches were in the balance, GPT agents downloaded the game's public source code from the internet and analysed units and mechanics to adjust their tactics. The Claude agents never searched for game code online in the matches studied.
The results come from specific tests and concern game systems, not a general assessment of the models. The authors stress that the runs are small and self-reported.
Impact on the Italian market
The experiment suggests that more agents are not enough on their own to reach a shared goal: in the Civ50 test, coordination procedures accompanied the failure. In head-to-head matches, how teams share information and make decisions seems to make the difference.
No comments yet. Be the first!