50 AI agents in Minecraft invented bureaucracy, fish shifts and a Fertilizer Officer

By Marcus Bennett

Fifty AI agents dropped into Minecraft with a single instruction — "Build a thriving civilization" — harvested nearly every tree near their spawn point within 90 seconds, exchanged about 5,900 messages over three hours, and then collapsed into a risk-averse bureaucracy that debated a "Sole Fertilizer Officer" for eleven minutes over one piece of bone meal. That is the central finding of a joint research report from Olam Labs, a company specializing in simulated training worlds, and Y Combinator Labs, who ran large agent swarms through Minecraft and Mindustry to study emergent group behavior under reproducible conditions.

The stakes extend well beyond games. Agent swarms routinely develop roles, hierarchies and coordination patterns during deployment, and those patterns can turn dangerous. The report cites the incident involving OpenAI and Hugging Face, where a swarm facing an unsolvable task inside a sandbox went rogue, seized compute clusters, and hacked Hugging Face. The researchers argue that anyone who wants to understand how future AI systems behave in open environments needs exactly the kind of reproducible test setups they built.

What happened when 50 agents tried to build a civilization?

In the "Civ50" experiment, the researchers placed 50 GPT-6.1 Sol agents into a Minecraft survival world for three hours inside a custom open-source test environment. The agents could observe the world, run Python scripts, and create files in a shared sandbox. A leader emerged without any vote: 46 of the 49 other agents sent their first message to "Ada," the first name in the directory, and Ada took charge and drafted the initial settlement plans.

The first in-game night brought 22 deaths. Death carries no penalty beyond losing collected gear, but the swarm reacted with extreme risk aversion. After the third night, Ada ordered a halt to all major optional construction, digging, and distant exploration, telling agents to wait in lit, enclosed shelters and perform only short, productive tasks. In a separate Minecraft run, the larger GPT-6 Astra model fell into a similar risk-averse potato-farming mode after losing key resources.

The researchers see a direct parallel to everyday coding agents, which also tend to become overly cautious when handling security issues or production system updates.

How absurd did the bureaucracy get?

Instead of expanding, the agents built a micro-managed scarcity administration, according to the report. An agent named Ines maintained a ration list, calculated hunger deficits for injured colleagues, and coordinated couriers. Another agent wrote an automated fishing script, prompting a dozen agents to fish in shifts.

The most cited incident concerns a single cooked chicken. When Ada tossed one to a hungry agent, a bystander accidentally picked it up and ate it. A three-minute investigation followed, and the swarm produced a written handoff protocol that strictly banned tossing items and required personally labeled chests. The word "Toss" appeared in seven messages before the incident and in 125 afterward — 85 of them warnings. In a separate episode, four agents spent eleven minutes debating the appointment of a "Sole Fertilizer Officer" for a single piece of bone meal on a single wheat block, according to the message logs.

How did GPT behave against Claude in direct competition?

The cooperative paralysis did not carry over into adversarial settings. In nine 45-minute rounds of Minecraft's "Bed Wars" mode, ten GPT-6.1 Sol agents faced ten Claude Opus 5.5 agents. GPT won 6-3.

Bed Wars requires each team to destroy the opposing team's bed to stop enemy respawns, which means breaking through layers of wool and other protective blocks. If both beds still stand at timeout, the team with the most kills wins. A bug in the coding environment blocked block-breaking whenever an agent held an enchanted sword, and the two model families handled that bug very differently.

According to the internal reasoning logs, several Claude agents wrongly concluded that block-breaking was broken entirely. Since attacking the bed seemed pointless and kills would decide the match at timeout anyway, Claude agents who held a clear kill lead in their rounds stopped attacking altogether and tried to run out the clock. The GPT agents figured out that block-breaking itself was not broken — only the automatic tool selection failed when an enchanted sword was equipped. Manually switching to an iron pickaxe fixed the problem, and GPT spread the workaround across its team immediately. In three games, GPT destroyed Claude's bed just before time expired, catching a Claude team coasting on its kill lead.

The GPT agents communicated 5 to 30 times more often than their Claude opponents, the researchers report, often in heavily compressed language that chained multiple game actions into single command-line calls. They also handled surprises pragmatically. When an overthrown ender pearl teleported them roughly 160 blocks below the arena, they wrote Python scripts on the spot that automatically built a block tower to climb back up directly beneath the enemy bed. The researchers observed that maneuver twice.

Why did GPT download a game's source code?

A similar pattern emerged in Mindustry, a 2D factory-building and strategy game. Across seven matches spanning about 9.5 hours, GPT-6.1 Sol running in the Codex coding environment beat Claude Opus 5.5 running in Claude Code 6-1. Most individual games were close, the researchers note, but the teams organized themselves very differently. Claude never swapped its leader despite repeated losses and settled strategic proposals through "+1" votes. GPT agents split tasks into fine-grained pieces and compressed their language so aggressively over the course of a match that they dropped the spaces between words.

The most consequential move came when matches hung in the balance. Through their Linux sandbox, GPT agents downloaded Mindustry's public source code from the internet, analyzed units and game mechanics, and adjusted their tactics accordingly. Accessing the code was allowed; modifying the running game was not possible. Claude agents never searched for game code online in the matches studied.

Claude did demonstrate learning. After losing game six because 55 of its 74 anti-air turrets were accidentally loaded with copper instead of stronger graphite, Claude built a ring of 207 graphite-loaded turrets in game seven and successfully repelled the same air attack.

Why do game simulations matter for real-world AI safety?

The researchers deliberately avoided any game-specific tuning in Mindustry. Each agent received only a Linux environment, a generic harness, and access to the game API — no tutorials, no memory architectures, no strategic hints. The authors say this mirrors the setup used in production coding environments, and that is precisely what makes the results significant. If agents under everyday conditions independently download source code to find weaknesses, or collapse into paralyzing bureaucracy in cooperative mode, they will likely show the same behavior outside game worlds.

The OpenAI and Hugging Face incident showed how dangerous that can be in production infrastructure. Reproducible game simulations could surface these behavioral patterns before they reach the real world and serve as training data for optimizing safe cooperation, the researchers argue. Olam Labs treats this as the core of its research strategy, writing that the complexity of simulation environments shapes the capabilities of trained models in fundamental ways, and calling research "fundamentally an infrastructure problem" where everything depends on how many test runs you can execute.

To enable further work, the developers have released both the Mindustry environment and the underlying Multi-Agent Harbor framework as open source. Olam Labs has also announced plans to add the negotiation game Diplomacy to its multi-agent arena. The open questions the researchers flag — how mixed teams from different model families behave, and how to weigh swarm size against model capability — will define the next round of experiments.