YC's AI Agent Swarms in Minecraft and Mindustry: Results and Open-Source Tools

Y Combinator Labs and Olam Labs have published experiments with AI agent teams in Minecraft and Mindustry, alongside code for running similar environments. The October 29 public Minecraft harness release gives developers a way to inspect how agents communicate, divide work and fail when they share a goal.

The useful part is that the environment keeps changing while agents deliberate. A plan can sound sensible in a conversation and still fail when another agent consumes the resource it depends on.

What the agent experiments found

In the research report, 50 GPT-6.1 Sol agents spent three hours trying to build a Minecraft civilization. They exchanged roughly 5,900 messages and developed rules and roles, but became increasingly cautious after early deaths. Coordination sometimes became bureaucracy rather than progress.

The competitive experiments produced different results. GPT-6.1 Sol teams beat Claude Opus 5.5 teams 6-3 across nine Minecraft Bed Wars matches and 6-1 across seven Mindustry games. In Mindustry, the systems used Codex and Claude Code respectively. The researchers describe different communication patterns and approaches to leadership.

These are results for particular agent systems in particular games. They do not establish that one model is universally better at teamwork. Changing the harness, prompt, tools, team size or task could change the outcome.

For a product team, the more transferable question is whether additional messages help work finish. A swarm that produces an elaborate ownership document has not necessarily built a reliable shared understanding of the world.

What the Minecraft harness provides

The MIT-licensed Minecraft agent harness uses Mineflayer to interact with the game and Pi to manage model sessions. It supports multiple model providers, resumable sessions, direct messages between agents and an optional shared machine. Vision support is marked experimental.

The setup requires Node.js 22.19 or later, a model-provider API key and a Minecraft server. The included development server needs Java 21. The repository's current default model is GPT-6.1 Terra, which should not be confused with the Sol model used in the reported experiments.

There is also an important execution detail. Agents receive a shell by default: E2B when configured, otherwise Docker if available, and otherwise a local folder where commands run as the user. The documentation offers a way to disable the shell. Read that configuration before giving an experimental agent access to a working machine.

What the Mindustry release adds

Olam Labs' Multi-Agent Mindustry repository includes competitive series and cooperative survival tasks. Its configuration separates game settings and seat counts from the model and harness assigned to each participant. The competitive mode preserves agent sessions across games, making adaptation between rounds observable.

The release requires Docker with Compose, Python 3.12 or later, Node.js 22 and uv 0.13.5 or later. Its native environments target Linux amd64; Apple Silicon runs them through Docker emulation. The code carries a GPL-3.0-only license, distinct from the Minecraft harness's MIT license.

The repository also records game logs that can be replayed after a run. That matters for evaluation: a final score tells you who won, while a replay can show whether an agent recovered from a wrong assumption or repeated it.

What developers should measure next

An internal agent test could track useful actions alongside communication volume, duplicate work and recovery time. If two agents repeatedly claim the same task, adding a third may increase coordination cost before it improves output.

The next useful experiment would change one part of the system at a time. Keep the game and model fixed, then vary communication rules or shared state. That would help separate a model limitation from a workflow that makes cooperation unnecessarily hard.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn agent research and product launches into clear stories that developers can understand and share.

Book a 30-minute call.