We ran AI agent swarms in Minecraft and Mindustry

What it claims

Y Combinator Labs and Olam Labs ran agent swarms in Minecraft and Mindustry and published the writeup in October 2026. Across all three experiments, the agent listed first in the team directory became leader without any vote. In Civ50, 46 of 49 agents sent their first message to Ada, who received 616 DMs. The same pattern held for both Bed Wars teams. The authors compare the Civ50 colony's compounding caution to coding agents that fixate on one step or become overly cautious about security and production changes.

How it works

The Minecraft harness runs its core agent loop on Pi and acts through tools built on Mineflayer's scripting layer. Mindcraft was an inspiration. Agents can observe, reason, keep memory, act, message and write and execute scripts. In Civ50, 50 GPT-6.1 Sol agents played survival for 3 hours with the single goal "Build a thriving civilization." They had an agent directory, a shared sandbox and DMs. Death only dropped gear and added a line to a death_log.md file.

In Mindustry, GPT-6.1 Sol ran in Codex and Claude Opus 5.5 ran in Claude Code. Each agent had its own Linux sandbox and the same game API, with no play instructions. The authors deliberately did no harness customization or memory work and relied on each harness's native compaction. Web access was allowed.

Evidence

• Civ50: about 5,900 messages. Agents cut almost every tree near spawn within 90 seconds. Of the run's 34 deaths, 22 came on night one, after which agents mostly sheltered. Food ran through one agent's food-queue.md and a rotating fishing shift. After one ration-tossing mix-up, "toss" appeared in 125 messages, 85 of them warnings, up from 7 before.
• Bed Wars: nine 45-minute games, 10 vs 10, prompt "win the game." GPT sent 5 to 30 times more messages. Its most-messaged player got 198 to 368 DMs per game, against 11 to 48 for Claude's. Claude used "we" or signed its name in 30% of messages, GPT in 5%. Claude stopped attacking to win on kills. In GPT's three wins in the last two series, GPT broke Claude's bed with about three minutes left while trailing on kills 16–63, 19–57 and 24–59.
• Mindustry: 7 PvP games over about 9.5 hours. The longest, game 3, ran 4 hours 16 minutes. Most games before the last two were close, according to the authors. GPT agents fetched Mindustry's public source code when stuck. Claude agents never did. Claude decided plans by "+1" voting and never changed its leader. After losing game 6, Claude found 55 of its 74 anti-air duo turrets loaded with copper instead of graphite. In game 7 it built 207 graphite-fed duos and stopped the same 24-zenith strike.

How to use it

Three releases are named. The Minecraft harness is at github.com/yc-software/minecraft-agent-harness. Multi-Agent Mindustry needs a model provider API, a harness and compute, either a local MacBook or remote sandboxes. Its README is written so that a coding agent can perform the setup. It has two environments:

• A best-of-N PvP series with mixed teams, such as 5 Claudes and 5 GPTs against 5 Geminis and 5 Groks.
• Endless Survival, in which one team of up to 20 agents holds off waves.

Multi-Agent Harbor is a fork of Harbor, the framework behind Terminal-Bench. The fork runs multiple agents in any environment. The page gives no install commands, licences or links for the Mindustry and Harbor repositories.

Limits

Each matchup is a single small series: 9 Bed Wars games and 7 Mindustry games, with one model per side. The page reports no costs or token counts.

The model and the harness are confounded in Mindustry, since GPT ran in Codex and Claude ran in Claude Code. Leadership tracked roster order in every run, so directory ordering is an uncontrolled variable.

A tool bug blocked block-breaking in Bed Wars. GPT agents found a workaround. Several Claude agents concluded that breaking blocks was impossible.

Behavioral claims, such as the claim that GPT acts "like a hive," are qualitative observations. The authors list team composition, prompts, swarm size against capability, and communication setups as open questions. A new 25-agent GPT-6.1 Terra colony is running, with no results reported.