YC Labs and Olam Labs ran 50 AI agents in a Minecraft civilization experiment

Fifty agents, one survival world

YC Labs, the research group inside Y Combinator, and Olam Labs have published a research report titled Agents in Games describing an experiment in which 50 autonomous agents were placed in a Minecraft survival world with the stated goal of building a lasting, prosperous settlement. According to ZDNET, which reviewed the report in detail on 2 November 2026, each agent ran on OpenAI's GPT-6.1 Sol and the session lasted three hours. The report itself was released on 29 October 2026.

The teams framed the exercise as an environment test rather than an accuracy benchmark: the point was to observe what happens when many models share a world, not to produce a leaderboard score. To make the run repeatable, they built a custom agent harness and released it as open source on GitHub, so other teams can rerun the same scenario with their own models.

What the colony actually did

The documented results mix survival statistics with organizational behavior. Of the 50 agents, 34 died over the session, and 22 of those deaths came during the first in-game night. Across the run the colony exchanged roughly 5,900 messages, and after the first-night losses the report describes a clear behavioral shift: agents became increasingly cautious, prioritizing survival and administration over expansion.

The more striking findings concern structure. Without any predefined social design, the agents created formal roles and permissions for themselves. The report's role list includes entries such as fishing shift, rod holder and hub steward. In one episode, four agents spent 11 minutes appointing a sole fertilizer operator responsible for applying a single piece of bone meal to a single wheat block. The episode is a compact illustration of how an open-ended game world turns into rule-bound organizational work once agents are handed goals, tools and a shared space.

Same harness, different models

The report also covers adversarial tests, including a Bed Wars match at 10-versus-10 scale in which agents running GPT-6.1 Sol faced agents running Claude Opus 5.5. According to the report, the GPT-based teams favored aggressive strategies and, after diagnosing a bug in one of their tools, passed workarounds to each other. Several Claude-based agents, by contrast, concluded that breaking blocks was impossible.

The authors present the comparison as descriptive rather than a ranking: it documents how two model families behave inside the same game harness, not which one wins. It also underlines a broader point of the report — harness design shapes outcomes as much as model choice, since a tool bug on one side and a mistaken belief about basic game mechanics on the other produced visibly different behavior in the same environment.

Why it matters

For developers building multi-agent systems, the experiment is a worked example of emergent organization. Coordination, role hierarchies and high communication volume appeared without anyone designing them, and because the harness is open source, the whole setup is reproducible: teams can rerun 50-agent sessions and compare models on coordination, tool diagnosis and role formation in the same scenario.

The numbers give the young practice of agent evaluation a concrete dataset — death counts, message volume, role inventories — rather than anecdote. Session length matters too: three hours is long enough for day-night cycles and resource pressure to shape behavior, which is what produced the documented death spike on the first night. And the Bed Wars results are a reminder that when agents fail, the explanation may lie in the harness rather than the model. A tool bug that one group diagnosed and routed around, and a false belief about block mechanics on the other, both left fingerprints on the outcome — evidence that environment design deserves as much scrutiny as model selection.