50-agent swarm fails Minecraft civilization test, falls into bureaucracy

Researchers at ycombinator and Olam Labs ran 50-agent swarms in Minecraft; the colony never thrived and agents turned to bureaucracy.

What happened

Tom's Hardware reports that AI researchers at ycombinator and Olam Labs recently tested AI agent swarms in gaming scenarios, using Minecraft and its various game modes as sandboxes where hundreds of agents were unleashed to work together or as adversaries. Per the report, the idea seems to have been that in-game collaborations and conflicts would provide insight into real-world cases involving AI agents. The source blog mentions two contrasting cases: thousand-agent swarms that "recently obliterated hundreds of open mathematical problems," and the "extremely dangerous" OpenAI-Hugging Face incident. The work builds on prior frameworks, specifically repos like Mindcraft and scripting tools in Mineflayer.

The first test was collaborative: 50 agents were tasked with building a thriving civilization in three hours. Tom's Hardware writes that it seems to have tested 50 GPT-6.1 Sol agents here, and not any Claude agents. In this "Civ50" run, the colony never really thrived. After the first night saw zombies take out a large number of agents, with little penalty, the researchers noted that "many ambitious initial tasks were abandoned." From then on the colonists cautiously survived, with a tendency toward bureaucracy. In one example, "four agents spent 11 minutes appointing a 'sole fertilizer operator' for a single piece of bone meal on a single wheat block." The researchers say the behavior "felt similar to patterns seen while working with coding agents."

The adversarial tests put 10 GPT-6.1 Sol agents against 10 Claude Opus 5.5 agents in Bed Wars. GPT agents came up with more aggressive and elaborate strategies, while Claude was more defensive. When a tool bug stopped block-breaking, GPT diagnosed it and shared workarounds, while several Claudes concluded breaking was impossible. In Mindustry, a light RTS combat game with tower defense mode, the GPTs won 6-1 across the 7 games.

Why it matters

The study puts numbers and a named scale on a multi-agent failure: 50 agents, a broad goal, a three-hour window, and a colony that never really thrived. The researchers themselves connect the pattern back to development work, saying the swarm behavior "felt similar to patterns seen while working with coding agents." The 11-minute debate over a "sole fertilizer operator" for one piece of bone meal is a concrete, citable instance of that behavior rather than an abstract warning.

The adversarial runs record differences between the two model families inside the same scenario. Tom's Hardware reports that GPT agents produced more aggressive and elaborate strategies and that Claude agents were more defensive, and that when a tool bug stopped block-breaking, GPT diagnosed it and shared workarounds while several Claudes concluded breaking was impossible. The researchers concluded that GPTs acted more like a hive, while Claudes had a single leader with a democratic-style decision process, and that in Mindustry the hive pattern was the foundation for a more consistently winning strategy.

For builders, the practical detail is that these swarms are open sourced, so the setups can be run, tweaked and tested directly rather than taken on description alone. The source blog ends with open questions and suggested directions, including comparing different mixes of models and varying sizes of agent swarms in Minecraft. The work also sits on existing public tooling, Mindcraft and Mineflayer, which is the layer anyone reproducing the runs would be working with.

What we don't know yet

Tom's Hardware hedges on the composition of the collaborative run, writing that it "seems" to have used 50 GPT-6.1 Sol agents and no Claude agents, so what was the exact roster? What criteria defined a "thriving civilization," and how was the Civ50 outcome scored? Neither the headline math result, the thousand-agent swarms said to have "obliterated hundreds of open mathematical problems," nor the "extremely dangerous" OpenAI-Hugging Face incident is described or independently verified in the source text. How many Bed Wars matches were played, and were the Mindustry and Bed Wars setups otherwise identical across teams? Has anyone outside ycombinator and Olam Labs rerun the open-sourced swarms and reported results?

Sources: tomshardware.com

How this Deep Take is made: it is written with AI assistance from the linked sources, then checked against them sentence by sentence, and every fact links to where it came from.