50 AI agents in Minecraft got more cautious the longer they ran, and in Bed Wars the GPT team beat the Claude team 6-3.
Civ50: 50 GPT-6.1 Sol agents spent 3 hours in survival Minecraft with the goal "Build a thriving civilization." They built a ration ledger, written rules and permits, and sent about 5,900 messages.
Deaths: 22 of the run's 34 deaths happened on the first night, and the agents became more risk-averse afterward. Ada became the de facto leader without any vote.
Bed Wars: Nine 45-minute games pitted 10 Claude Opus 5.5 agents against 10 GPT 6.1 Sol agents. GPT usually broke Claude's bed in the final minutes, while Claude played defensively and aimed to win on kills.
Open source: The team released the Minecraft agent harness they built for the experiment on GitHub.
Sign in to comment.
TL;DR of the article, generated automatically.
50 AI agents in Minecraft got more cautious the longer they ran, and in Bed Wars the GPT team beat the Claude team 6-3.