# AI agents in Minecraft

> A 50-agent survival experiment.

By BIGCHANGE.AI Editorial

Published: 2026-11-02T13:18:39.484Z
Updated: 2026-11-02T13:18:39.484Z
Canonical: https://bigchange.ai/blog/yc-minecraft-agents

![Conceptual charcoal illustration of several orange paths winding across a stepped block terrain and meeting partway up.](https://bigchange.ai/api/media/file/gemini4-minecraft-coordination-terrain-hero-v2.png)
AI-generated conceptual illustration by BIG CHANGE.

YC Labs and Olam Labs report a 50-agent Minecraft survival experiment. The October 2026 report was [public by October 29](https://ax-news.com/digest/2026-10-29).

## The big change

- Agents coordinated around food and shelter, leaving records of their actions.
- Communication was observable, but one run cannot establish a general rule.
- The released tools enable new tests; code is not independent replication.

## What happened in the settlement

The [researchers](https://labs.ycombinator.com/projects/games/index.html) gave 50 GPT-6.1 Sol agents a civilization-building goal, a directory, direct messages and shared sandbox. They report about 5,900 messages and 34 deaths in three hours, including 22 on the first night. Agents dropped gear on death and respawned.

The group kept a ration ledger; Ada, first in the directory, became a coordinator. The authors interpret sheltering and fewer risky tasks as caution, not evidence of fear.

## Tools for further tests

The [Minecraft harness repository](https://github.com/yc-software/minecraft-agent-harness) describes persistent model sessions, Mineflayer game tools, messaging and shell sandboxes. A harness connects a model to observations and actions; changing the tools or communication available can change what a test measures. The repository allows others to inspect those choices.

BIG CHANGE has not run it or reproduced the settlement.

Separate research has explored variation within Minecraft. The authors of [TeamCraft](https://arxiv.org/abs/2412.05255) describe a benchmark with many task variants and report difficulty generalizing to unseen goals, scenes and numbers of agents. 

That separate study does not replicate the settlement run.

A [previous BIG CHANGE report](https://bigchange.ai/blog/gemini-4-agents-minecraft-27-hour-experiment) covered a four-agent Minecraft completion attempt.

## Sources & further reading

- [YC Labs report](https://labs.ycombinator.com/projects/games/index.html): the settlement setup and reported observations.
- [Minecraft harness](https://github.com/yc-software/minecraft-agent-harness): code and configuration.
- [TeamCraft paper](https://arxiv.org/abs/2412.05255): separate benchmark and generalization study.
- [October 29 discussion archive](https://ax-news.com/digest/2026-10-29): evidence the report was public by then.

## Sources

- [YC Labs: Agents in Games](https://labs.ycombinator.com/projects/games/index.html) — Settlement setup and reported observations.
- [Minecraft agent harness](https://github.com/yc-software/minecraft-agent-harness) — Minecraft agent code and configuration.
- [TeamCraft paper](https://arxiv.org/abs/2412.05255) — Separate benchmark and generalization study.
- [October 29 discussion archive](https://ax-news.com/digest/2026-10-29) — Evidence the report was public by then.