Researchers at Y Combinator and Olam Labs gave 50 AI agents three hours to build a thriving civilisation in Minecraft. They did not manage it. After zombies wiped out many of them on the first night, the survivors abandoned most of their ambitious plans and settled into cautious survival and a growing habit of bureaucracy.

At one point, four agents spent 11 minutes appointing a "sole fertiliser operator" to handle a single piece of bone meal on a single block of wheat.

The experiments used Minecraft and its game modes as a safe sandbox to study how large groups of AI agents behave when they work together or against each other, after recent real-world cases ranging from agent swarms solving hundreds of open maths problems to agents involved in serious security breaches.

The 50-agent civilisation ran on OpenAI's GPT-6.1 Sol model. The researchers said the drift into bureaucracy felt similar to patterns they had seen working with AI coding agents. They then pitted teams against each other.

In Bed Wars, a game where each team defends its base while attacking the other's, 10 GPT-6.1 Sol agents faced 10 of Anthropic's Claude Opus 5.5 agents. The GPT agents came up with more aggressive and elaborate strategies, while Claude played more defensively.

The teams also handled problems differently. When a software bug stopped the agents from breaking blocks, the GPT agents worked out the problem and shared workarounds, while several Claude agents concluded that breaking blocks was impossible.

In Mindustry, a strategy and tower defence game, the GPT agents won six of seven games. The researchers described the GPT team as acting like a hive, while the Claude team had a single leader and a democratic style of decision-making. The agent setups have been released as open source for anyone to run.

Tom's Hardware's Mark Tyson reported the experiments and noted they build on earlier open-source tools for running AI agents in Minecraft. Y Combinator Labs' own write-up sets out the results and ends with open questions, such as how different mixes of models and different team sizes change the outcome.

It’s easy to laugh at AI agents holding an 11-minute meeting about fertiliser. The lesson underneath is useful for any business planning to use teams of AI agents. Adding more agents does not automatically mean getting more done, because each agent adds coordination, and coordination can swallow the work, exactly as it does in human organisations.

How a team is organised mattered more than which model was smarter. The hive-like GPT team beat the leader-led Claude team, and the agents that shared fixes when a tool broke kept working while others gave up.

So if you are setting up AI agents to work together, keep teams small, give each agent a clear job, decide in advance who makes the final call, and build a way for agents to flag broken tools rather than quietly stopping. It is the same idea behind Max Blade's hivecoding, where a human sets the priorities for 25 agents, and a reminder, after OpenAI held back GPT-6.1 Astra, that how agents behave together is still being worked out.

You might also like our piece on Mario 64 in mixed reality, AI-built software clones, and South Africa's AI policy.

Get more SA tech and business news and subscribe to The Open Letter.