What changed
AI researchers at ycombinator and Olam Labs recently tested AI agent swarms in gaming scenarios, using Minecraft as a sandbox. They deployed 50 GPT-6.1 Sol agents to build a thriving civilization within three hours. The experiment revealed that the agents failed to establish a successful colony, instead falling into bureaucratic inefficiencies.
What this means in practice
The study highlights challenges in coordinating large AI agent swarms. After the first night, zombies eliminated a significant number of agents with minimal penalty, leading to abandoned tasks. The agents then focused on survival, with one example showing four agents spending 11 minutes appointing a 'sole fertilizer operator' for a single wheat block. Researchers noted this behavior mirrored patterns seen in coding agent collaborations.
Limitations and availability
The study also tested AI agents in conflict scenarios, such as Bed Wars and Mindustry. In Mindustry, GPT agents won 6-1 against Claude Opus 5.5 agents, with GPTs acting more like a hive while Claudes used a democratic decision process. The researchers concluded that GPTs formed a more consistent winning strategy. The project is open-sourced, allowing others to run and tweak the experiments. The study raises open questions about model mixtures and swarm sizes, encouraging further investigation.
