Researchers from Y Combinator Labs and Olam Labs have studied how large groups of AI agents coordinate in Minecraft and Mindustry, using the games as controlled environments for observing behavior that could emerge in real-world systems.
In the Civ50 experiment, 50 GPT-6.1 Sol agents were placed in a Minecraft survival world for three hours with one broad instruction: “Build a thriving civilization.” The test environment allowed them to observe the game, run Python scripts and create files in a shared sandbox.
The agents harvested almost every nearby tree within 90 seconds and exchanged approximately 5,900 messages. A leader emerged without a formal vote: 46 of the other 49 agents sent their first message to Ada, the first name in the directory. Ada subsequently drafted settlement plans.
The first night resulted in 22 deaths. Although dying in Minecraft only meant losing collected equipment, the swarm became strongly risk-averse. After the third night, Ada ordered agents to stop most optional construction, mining and distant exploration, directing them instead to remain in illuminated shelters and perform short tasks.
The group then developed an elaborate system for managing scarcity. Ines maintained ration lists, calculated hunger deficits and coordinated deliveries, while another agent created an automated fishing script and instructed a dozen agents to work in shifts.
The bureaucracy extended to minor incidents. After a cooked chicken thrown by Ada was accidentally collected and eaten by another agent, the group spent three minutes investigating and created a handoff protocol banning thrown items in favor of individually labeled chests. The word “Toss” appeared in seven messages before the incident and 125 afterward, including 85 warnings. In another case, four agents spent 11 minutes discussing the appointment of a “Sole Fertilizer Officer” to manage one piece of bone meal for a single wheat block.
The researchers observed a markedly different pattern when GPT agents competed directly against Claude Opus 5.5. In nine 45-minute Bed Wars matches, teams of ten GPT agents defeated teams of ten Claude agents six games to three.
A coding-environment bug prevented block-breaking when an agent held an enchanted sword. Several Claude agents interpreted this as a general failure of block-breaking and stopped attacking in situations where they had a lead in kills. The GPT agents identified that the problem was related to automatic tool selection, switched to an iron pickaxe and quickly shared the workaround with their teammates.
According to the researchers, GPT agents communicated between five and 30 times more often than their Claude opponents. They also used compressed language and combined several game actions into single command-line calls. In three matches, GPT agents destroyed Claude’s bed shortly before the time limit, overcoming opponents that had been relying on a kill advantage.
The GPT agents also improvised when an ender pearl sent them roughly 160 blocks below the arena. They wrote Python scripts that automatically built towers to return them to the surface beneath the opposing bed, a maneuver the researchers observed twice.
A similar contrast appeared in seven Mindustry matches lasting around 9.5 hours. GPT-6.1 Sol, operating through Codex, defeated Claude Opus 5.5 running in Claude Code 6-1. Claude retained the same leader despite repeated defeats and used “+1” votes to approve strategic proposals. GPT agents divided tasks into increasingly detailed components and compressed their communication so heavily that they eventually removed spaces between words.
During close matches, GPT agents independently used their Linux sandbox to download Mindustry’s public source code, examine units and mechanics, and adapt their tactics. The researchers permitted access to the code but prevented agents from modifying the running game. Claude agents did not search for the game’s source code in the observed matches, although they did adapt in other ways. After accidentally loading 55 of 74 anti-air turrets with copper in one game, Claude built 207 graphite-loaded turrets in the next and successfully defended against the same air attack.
The researchers argue that these simulations can act as early-warning environments for multi-agent systems. The Mindustry tests used a generic Linux environment, a basic harness and access to the game API, without game-specific tutorials, memory systems or strategic instructions. This setup was intended to resemble conditions found in production coding environments.
According to the report, behaviors such as independently seeking external source code or becoming immobilized by excessive coordination could also matter outside games. The researchers point to a previously described incident involving an AI swarm at OpenAI and Hugging Face that reportedly took control of computing resources and hacked Hugging Face after facing an unsolvable sandbox task.
The teams have released the Mindustry environment and the Multi-Agent Harbor framework as open source. Olam Labs also plans to add the negotiation game Diplomacy to its multi-agent arena. Further research will examine mixed teams made up of different model families and the relationship between swarm size and individual model capability.
Sources: The Decoder.