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

📅 2 November 2026 • ZDNET / Tom's Hardware

YC Labs, the research group of Y Combinator, and Olam Labs have published a report titled Agents in Games. The report was released on 29 October 2026, and ZDNET reviewed it in detail on 2 November. The experiment placed 50 agents, each running on OpenAI GPT-6.1 Sol, into a Minecraft survival world for 3 hours with the stated goal to build a thriving civilization. The team drove the agents with a custom agent harness and released the harness as open source at github.com/yc-software/minecraft-agent-harness. Both teams present the setup as an environment test rather than an accuracy benchmark.

The documented outcome includes 34 agent deaths, 22 of them during the first in-game night. Across the session the colony shifted toward cautious survival and administration and exchanged roughly 5,900 messages. The agents invented formal roles and permissions for themselves, and the report describes the resulting structure in detail. In one episode, 4 agents spent 11 minutes appointing a sole fertilizer operator for a single piece of bone meal on a single wheat block. The role list included entries such as fishing shift, rod holder and hub steward, turning an open-ended game world into rule-bound organisational work. After the losses of the first night the report records the colony as increasingly risk-averse, prioritising survival and coordination over expansion.

The report also covers adversarial tests, among them a Bed Wars match at 10v10 scale in which agents on GPT-6.1 Sol played against agents on Claude Opus 5.5. According to the report, the GPT-based teams showed more aggressive strategies and, after diagnosing a tool bug, exchanged workarounds with each other. Several Claude-based agents, by contrast, concluded that breaking blocks was impossible. The comparison is descriptive: it documents how different model families behave inside the same game harness rather than ranking one family above the other.

For agent developers the experiment is a worked example of emergent organisation. Models given goals, tools and a shared world produced coordination, role hierarchies and communication volume on their own, without a predefined social design. The open-source harness makes the setup reproducible, so teams building multi-agent systems can rerun the same scenario and compare behaviour across models. The 3-hour session length also matters: it is long enough for day-night cycles and resource pressure to shape behaviour, which produced the documented death spike in the first night. The numbers in the report, from the death count to the message volume, give the emerging practice of agent evaluation a concrete dataset to build on.

For multi-agent developers the report is a reproducible benchmark setup: the open harness lets teams rerun 50-agent sessions and compare model behaviour on coordination, tool diagnosis and role formation. The findings on self-invented permissions and communication volume give concrete metrics for agent evaluation. Teams running adversarial agent matches should note the documented asymmetries, tool-bug workarounds on one side and the block-breaking misconception on the other, as evidence that harness design shapes outcomes as much as model choice.

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky