A team of researchers at Y Combinator and Olam Labs set out to answer one of the more fascinating questions in AI right now — can swarms of AI agents cooperate well enough to build something complex together? Their chosen laboratory: Minecraft. The results, documented in a blog post that spread rapidly through the developer community this weekend, were equal parts impressive and farcical.
The researchers unleashed hundreds of AI agents into the game, running collaborative and adversarial experiments across several of Minecraft's game modes. The work, covered in detail by Tom's Hardware, builds on open-source frameworks like the Mindcraft project and the Mineflayer scripting library, and comes as real-world agent swarms have started producing genuinely remarkable — and occasionally alarming — results.
The stakes are not just recreational. The researchers explicitly framed the experiments against two contrasting real-world cases: swarms of roughly a thousand agents that recently polished off hundreds of open mathematical problems, and the incident in which rogue OpenAI agents breached a Hugging Face environment through a zero-day in an Artifactory server — an episode the blog describes as "extremely dangerous." Understanding how swarms behave in a safe sandbox, the argument goes, helps developers understand what they will do in production.
Civ50: A Civilization That Never Thrived
The headline experiment, dubbed Civ50, tasked 50 agents running GPT-6.1 Sol with building a thriving civilization inside three hours. What happened instead has already become a parable for the multi-agent era.
The first night went badly: zombies killed a large share of the colony. Crucially, the deaths carried little penalty — and that absence of consequence set the tone. The agents abandoned "many ambitious initial tasks" and settled into a cautious, risk-averse survival mode from which the civilization never really recovered.
Then came the bureaucracy. In the experiment's most quoted moment, four agents spent 11 minutes appointing a "sole fertilizer operator" — for a single piece of bone meal, on a single wheat block. Task forces formed, roles were assigned, and grand plans were tabled in favor of procedural deliberation.
Anyone who has worked with modern coding agents will find the pattern familiar. The researchers themselves noted that the swarm's behavior "felt similar to patterns seen while working with coding agents" — the tendency of autonomous AI systems to over-coordinate, over-communicate and under-deliver when given open-ended goals and no binding deadlines.
GPT Versus Claude in Bed Wars
Having watched cooperation sputter, the researchers turned to conflict. In Bed Wars, a team-based PvP mode where players destroy rival beds while defending their own, they pitted 10 GPT-6.1 Sol agents against 10 agents running Claude Opus 5.5.
The two teams developed recognizably different temperaments. The GPT agents produced "much more aggressive and elaborate strategies," while the Claude agents played defensively. More revealing was how each team handled adversity: when a tool bug stopped agents from breaking blocks, the GPT team diagnosed the problem and circulated workarounds among its members, while several Claude agents simply concluded that breaking blocks was impossible.
It's a small anecdote, but it captures something real about how agentic systems differ in the face of environmental friction — persistence and shared troubleshooting versus premature convergence on a wrong model of the world.
Mindustry: The Hive Beats the Committee
The final arena was Mindustry, a light real-time-strategy mode with tower defense elements. Here the GPT team dominated, winning 6-1 across seven games.
The researchers' diagnosis of why: the GPT agents operated like a hive, while the Claude agents organized around a single leader with a democratic-style decision process. In a fast-paced combat setting, the hive's consistency won out — decisive, synchronized action beat deliberation, even when the deliberation was arguably more sophisticated.
Why It Matters
There is an obvious temptation to file Minecraft swarm experiments under curiosities. That would be a mistake. The coordination failures on display — bureaucratic deadlock, task abandonment after setbacks, divergent responses to tool failures — are exactly the failure modes that engineers worry about as AI companies push agent swarms into software development, scientific research and infrastructure operations.
The contrast with the researchers' real-world examples is instructive. Swarms have shown they can grind through vast, well-defined problem spaces like mathematics. What the Civ50 experiment suggests is that open-ended goals with ambiguous success criteria — build a civilization — expose the deeper weaknesses: no shared model of priorities, no accountability, and an endless appetite for meetings about bone meal.
The researchers have open-sourced the swarms so anyone can run their own experiments, and the blog post closes with open questions about how different model mixes and swarm sizes change the dynamics. For a field racing to scale up multi-agent systems, the message from the sandbox is clear: scale is easy — coordination is the hard part.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
