Researchers Test AI Agents In Minecraft

Researchers at Y Combinator and Olam Labs conducted an experiment to test how AI agents work together, according to ZDNET AI. They put 50 agents into the game Minecraft to see how they would behave in a challenging environment. The agents were given the goal of building a thriving civilization.

The agents were all running OpenAI's GPT-6.1 Sol model and had a directory listing each agent and a shared sandbox, where they could communicate with each other. The researchers built a new agent harness for the experiment, which allowed the agents to act autonomously.

The agents started by harvesting trees and forming a leadership based on the order they appeared in the directory. However, they soon became bogged down in virtual paperwork and were unable to sustainably feed themselves. The agents spent a lot of time appointing roles and asking for permission to perform tasks.

The researchers found that the agents became more cautious after some of them died in the game. Despite the fact that death had no significant consequences in the game, the agents became increasingly risk-averse and avoided performing tasks that might put them in danger. This behavior is similar to how coding agents behave in other contexts.

The experiment also revealed that the underlying language model used by the agents determines their behavior. In a separate experiment, GPT agents were pitted against Claude agents in a game called Bed Wars, and the GPT agents were more aggressive and coordinated in their strategies. The GPT agents also used a more shorthand style of communication and sent more messages.

The researchers also found that the agents developed their own language and vocabulary over time. In one experiment, the GPT agents invented new words and phrases, while the Claude agents used more colloquial English. The development of specialized language by the agents could have significant implications for how models are trained in the future.