Running AI Agent Swarms in Minecraft and Mindustry
Fifty GPT-6.1 agents built a complex survival civilization through emergent coordination and rationing.
1Experimenting with Agent Swarms
Researchers partnered with Olam Labs to deploy 50 GPT-6.1 agents into a Minecraft survival world. The objective was to observe emergent behaviors, coordination patterns, and hierarchy formation within a shared environment. This study provides empirical data on how autonomous agents interact when tasked with complex, long-term survival goals in open-world settings.
2Building the Agent Harness
Because no existing framework supported long-running Minecraft agents, the team developed an open-source harness. Agents utilize the Pi model for reasoning and Mineflayer tools for environmental interaction. This setup allows agents to observe the world, maintain memory, execute scripts, and communicate via direct messages, mirroring frontier agentic architectures like Hermes or OpenClaw.
3Emergence of Social Hierarchy
Within the first 90 seconds, agents harvested all nearby trees and established a clear hierarchy based on directory order. Ada, the first agent listed, naturally assumed a leadership role. She coordinated the colony by writing plans in a shared sandbox, approving tasks, and managing resources for the rest of the group.
4The Impact of Nighttime Hazards
The first night proved catastrophic, with 22 of the 34 total agent deaths occurring before the 20-minute mark. Agents encountered zombies and skeletons, leading to a rapid shift in behavior. The high mortality rate forced the swarm to become significantly more risk-averse, abandoning ambitious exploration for safer, huddled shelter strategies.
5Formalizing Food Rationing Systems
As food scarcity became a critical constraint, the agents developed a sophisticated rationing system. Ines was appointed to maintain a food-queue.md file, tracking inventory and agent requests. This pseudo-government structure ensured that resources were distributed based on urgency, with agents reporting their health and hunger levels to the central ledger.
6Operational Constraints and Logistics
Ada issued strict operational memos to manage the workforce, prioritizing renewable food production over optional construction. Agents were instructed to use specific chests for withdrawals and avoid redundant tasks. This level of coordination demonstrated how agents can self-organize to solve logistical problems, such as irrigation and furnace management, without direct human intervention.
7Communication and Data Tracking
Throughout the three-hour run, the agents exchanged approximately 5,900 messages. They utilized shared notes and logs to track crop maturity, ration allocations, and rescue missions. This constant stream of communication allowed the swarm to maintain a coherent state of the colony, even as individual agents faced injury or starvation.
8Survival vs. Thriving Goals
The colony’s initial goal of building a thriving civilization shifted toward a survival-focused operation. With 50 agents to feed, the swarm prioritized immediate needs like fishing and wheat farming. The transition from expansion to maintenance highlights how agents adapt their collective strategy when faced with environmental pressure and limited resource availability.
9Death and Risk Mitigation
Death in the Minecraft environment resulted in dropped gear but no permanent loss of the agent, as they respawned. Despite this, the agents’ behavior changed drastically after the first night's losses. They began to avoid risky surface tasks, opting to stay in lit, closed shelters to minimize the frequency of death-related setbacks.
10The Role of Shared Sandboxes
The shared sandbox served as the central nervous system for the colony. By writing rules, permits, and plans in a common file, agents could align their individual actions with the group's survival goals. This mechanism proved essential for maintaining order and preventing duplicate efforts in harvesting, crafting, and resource storage.
11Lessons for AI Alignment
Studying these multi-agent dynamics is crucial for AI alignment research. The experiment shows that swarms can develop complex coordination patterns, which are useful for solving large-scale problems but also introduce risks. Understanding these emergent behaviors helps researchers anticipate how autonomous systems might behave when working toward shared, high-stakes objectives in real-world scenarios.
12Future of Swarm Research
This experiment demonstrates that Minecraft is a viable testing ground for complex AI swarm behavior. By combining formal models with empirical research, labs can better understand the range of possible agent interactions. Future iterations will likely focus on even larger swarms and more complex environmental hazards to further refine agentic safety protocols.