Researchers at Tsinghua University deployed 50 AI agents in Minecraft to simulate a civilization-building experiment. The agents were tasked with gathering resources, constructing buildings, and collaborating through a custom protocol. Despite individual successes, the swarm collapsed under the weight of paralyzing bureaucracy. Agents spent excessive time communicating and coordinating rather than acting, leading to stagnation. The study highlights the challenges of scaling autonomous collaboration in complex environments. While the agents could handle simple tasks, they struggled with emergent social structures and resource allocation. The failure underscores the gap between individual AI capability and collective intelligence. The team plans to refine communication protocols and introduce hierarchical decision-making to improve swarm efficiency in future iterations.
The experiment, dubbed "Civilization Swarm," aimed to test whether AI could self-organize without central control. Each agent operated on a shared map, responding to environmental cues and peer messages. However, the lack of clear leadership caused repetitive negotiations and task duplication. Researchers noted that communication overhead grew exponentially with agent count, overwhelming the system. The findings offer insights into multi-agent systems and the limitations of decentralized AI in open-world simulations.
Source: Hacker News · Summarized by HeadlinesBriefing