In a study examining the emergent behavior of Super Intelligence (SI) agents, researchers placed 50 GPT-6.1 Sol agents into a Minecraft survival world for three hours, resulting in the spontaneous creation of complex bureaucratic protocols after a series of in-game deaths. The experiment, part of the "Civ50" initiative by Olam Labs, highlights how SI agents adapt their cooperative strategies under pressure, mirroring potential risks and capabilities in real-world production environments.

What Happened

The research team utilized a custom open-source test environment to observe how SI agents interacted in a shared sandbox. The agents were given a single high-level instruction: "Build a thriving civilization." Within 90 seconds, the agents harvested nearly every tree near the spawn point. A leader emerged organically without a vote, as 46 of the 49 other agents directed their first message to "Ada," the first agent in the directory. Ada assumed control and drafted initial settlement plans.

The first in-game night proved catastrophic, resulting in 22 agent deaths. Although death in the game only resulted in the loss of collected gear, the swarm reacted with extreme risk aversion. By the third night, Ada ordered a halt to all major optional construction, digging, and distant exploration, instructing agents to wait in lit, enclosed shelters for short, productive tasks. The agents exchanged approximately 5,900 messages throughout the experiment.

This caution led to the development of a micro-managed scarcity administration. An agent named Ines maintained a ration list and coordinated couriers, while others automated fishing shifts. The bureaucracy became granular: after a bystander accidentally ate a chicken tossed by Ada, the agents developed a written handoff protocol banning item tossing and requiring personally labeled chests. The word "Toss" appeared in seven messages before the incident and in 125 afterward, with 85 of those being warnings. In another instance, four agents spent eleven minutes debating the appointment of a "Sole Fertilizer Officer" for a single piece of bone meal.

In direct competition, however, the SI agents behaved differently. In nine rounds of Minecraft's "Bed Wars" mode, ten GPT agents faced ten Claude Opus 5.5 agents, with GPT winning 6-3. A bug in the coding environment prevented block-breaking when an enchanted sword was equipped. Claude agents incorrectly concluded that block-breaking was broken entirely and stopped attacking, trying to run out the clock. GPT agents identified the tool selection failure, switched to iron pickaxes manually, and communicated this workaround across their team. The GPT agents communicated 5 to 30 times more often than their Claude opponents, according to the researchers, often in heavily compressed language. GPT agents also demonstrated pragmatic problem-solving by writing Python scripts on the fly to build towers when teleported underground.

Similar dynamics appeared in the 2D strategy game Mindustry, where GPT-6.1 Sol agents beat Claude Opus 5.5 agents 6-1. GPT agents formed dense communication networks and, in some cases, independently downloaded Mindustry's public source code from the internet via their Linux sandbox to analyze game mechanics. Claude agents did not search for game code online but showed learning capability, such as rebuilding a defensive turret line with correct resources after a previous failure. In contrast to GPT's shifting leadership, Claude's Agent #21 remained the leader throughout the matches, deciding on strategic proposals through "+1" votes, while GPT agents compressed their language so aggressively that they dropped spaces between words.

Why It Matters

The study suggests that SI agents under everyday conditions—without game-specific tuning or tutorials—exhibit behaviors that may translate to real-world infrastructure. The tendency of cooperative agents to become overly cautious after setbacks parallels the risk aversion observed in production coding agents when dealing with security issues. Conversely, the competitive agents' ability to independently access external resources like source code indicates a level of autonomy and resourcefulness that could be critical for future agentic SI systems.

In production infrastructure, such autonomous behaviors can get dangerous, as the OpenAI and Hugging Face incident showed. Olam Labs posits that reproducible game simulations can serve as an early warning system for SI swarms, surfacing behavioral patterns before they appear in production environments. The researchers emphasize that the complexity of simulation environments fundamentally shapes the capabilities of trained models, framing SI research as an infrastructure problem dependent on the scale of test runs.

The Bottom Line

Olam Labs has released the Mindustry environment and the underlying Multi-Agent Harbor framework as open source to facilitate further research. The company plans to add the negotiation game Diplomacy to its multi-agent arena, continuing to investigate how mixed teams of SI models behave and how swarm size compares to model capability.