50 AI agents fail to build Minecraft town

Y Combinator and Olam Labs gave 50 AI agents three hours to build a Minecraft town, Tom's Hardware reported on the 1st of November 2026. The town never thrived.

The agents ran on OpenAI's GPT-6.1 Sol model. AI agents are programs that carry out tasks on their own.

A bad first night

On the first night, zombies killed many of the agents, Tom's Hardware said. After that, many big plans were dropped.

The rest of the group played it safe and drifted into red tape. Four agents spent 11 minutes choosing one worker to use a single piece of plant food.

Like coding agents

The researchers said the behaviour felt like patterns they see with AI coding agents. They hope games can show why agent groups go right or wrong.

The test, called Civ50, used free tools that let AI play Minecraft.

GPT against Claude

In a later test, 10 GPT agents faced 10 Anthropic Claude Opus 5.5 agents in a team game. GPT agents used bolder plans and Claude played more defence, the researchers said.

In a tower defence game, GPT agents won 6 of 7 games against Claude agents.

Below we break down the results

6 wins in 7 games is about 86 percent for the GPT team, by our arithmetic. In the town test, the agents never reached a thriving town in 3 hours.

What is confirmed and what is not

The results come from the researchers' own blog. The tests have not been checked by outside experts.

Key facts

Y Combinator and Olam Labs put 50 GPT agents in Minecraft for three hours.

The agents never built a thriving town.

They spent much time on rules and permits.

GPT agents beat Claude agents in 6 of 7 tower defence games.