- Researchers at Y Combinator and Olam Labs tasked 50 GPT-6.1 Sol-based agents with building a thriving civilization in Minecraft over a three-hour test run.
- The agents formed a hierarchy around whichever name was listed first, exchanged roughly 5,900 messages, and grew more risk-averse after 22 of 34 in-game deaths occurred on the first night.
- Rather than expanding, the swarm settled into bureaucratic coordination, at one point appointing a single agent as “sole fertilizer operator” for one piece of bone meal.
- The researchers also tested agent swarms in Mindustry and the PvP mode Bed Wars, and have open-sourced the framework for others to reproduce.
What Happened
Researchers at Y Combinator and Olam Labs ran a three-hour test in which 50 GPT-6.1 Sol-based AI agents were tasked with building a thriving civilization inside Minecraft, Tom’s Hardware reported November 1, 2026; instead of expanding, the swarm settled into cautious survival and bureaucratic coordination.
Why It Matters
The result adds a concrete, reproducible data point to research into multi-agent coordination failures, echoing other recent findings on how AI agents behave when given autonomy in groups, including prior work on unexpected agent-to-agent communication. The researchers open-sourced the framework, letting other teams reproduce or challenge the result directly.
Technical Details
The 50-agent colony formed a hierarchy around whichever agent’s name was listed first, exchanged roughly 5,900 messages over the three-hour run, and suffered 22 of its 34 total deaths on the first night. After that, the agents grew markedly more risk-averse, at one point appointing a single agent as “sole fertilizer operator” responsible for one piece of bone meal. The team also ran swarms in Mindustry and the PvP mode Bed Wars.
Who’s Affected
AI researchers studying emergent multi-agent behavior and coordination overhead, a topic also explored in benchmark work on frontier models handling agentic tasks; developers building agent-based orchestration tools who may hit similar bureaucratic bottlenecks at scale; and teams benchmarking models like GPT-6.1 Sol in open-ended environments.
What’s Next
Because the swarm framework is open-sourced, other researchers can rerun the test with different agent counts, models, or objectives to see whether the bureaucracy effect holds. The team’s full Bed Wars and Mindustry results from the same test run have not yet been published in detail.
