IT Home On November 2, AI researchers from Y Combinator and Olam Labs recently tested an agent cluster in a game scenario. The study used “Minecraft” and various gameplay modes as test sandboxes, allowing hundreds of agents to collaborate or compete with each other. The goal was to use the collaboration and conflicts within the game to help people understand the agent events that cause widespread attention in the real world.
IT Home notes that this research helps clarify the advantages and disadvantages in recent cluster events of intelligent agents, making it easier for technology developers to optimize related capabilities. The blog post mentions two contrasting cases: one involves a cluster of over a thousand intelligent agents that “recently solved hundreds of public mathematical problems”; the other is the “high-risk” incident between OpenAI and Hugging Face. This research is based on previous achievements and frameworks, mainly relying on code repositories such as Mindcraft and script tools like Mineflayer.
Multi-agent interaction in "My World"
The first AI cluster test is a simple collaborative task. In this test phase, researchers arranged 50 AI agents to establish a thriving civilization community within three hours. This test used 50 GPT-6.1 Sol AI agents, without invoking any Claude series agents.
The conclusion that is most worth noting from this test called Civ50 is perhaps this: the community consisting of 50 agents never truly developed or grew. On the first night, a large number of agents were killed by zombies (but the punishment mechanism had a mild effect), and research records indicate that “many ambitious tasks set up at the beginning were left unfinished”.
After that, the group of colonists could only carefully maintain their survival, and a bureaucratic-like behavior tendency also emerged. An example is given in the text: “Four agents spent 11 minutes specifically appointing a ‘single fertilizer operator,’ yet the task was merely to use bone powder for a wheat field.” The researchers said that the behavior pattern of such agent clusters is very similar to what they observed in their programming agent project.
Next, the researchers arranged a cluster of agents to conduct adversarial testing. First, the “Waking Up War” mode was selected, which is a team PvP gameplay where the focus is on destroying the opponent’s base and defending one’s own base. The testing group consisted of 10 GPT-6.1 Sol agents against 10 Claude Opus 5.5 agents.
Researchers recorded: “The offensive strategies conceived by the GPT agent are more aggressive and complex, while the Claude agent tends to focus on defense.” The way both teams handle problems also shows interesting differences. “Due to a tool malfunction in the block destruction function, the GPT agent completed the problem diagnosis and shared a temporary solution; several Claude agents directly determined that the destruction operation could not be performed.”
Finally, the research team conducted tests in Mindustry, a lightweight real-time strategy module with a tower defense gameplay style, running within Minecraft. In seven matches, the GPT agent defeated the Claude agent with six wins and one loss—a good result for this OpenAI-developed agent. The researchers concluded that the GPT agent is more like a unified swarm, while the Claude agent has a single leader and adopts a democratic decision-making process. In the Mindustry scenario, the swarm mode is easier to develop a stable winning strategy.
This intelligent agent cluster project has been opened source. If you are interested, you can adjust the parameters and conduct replication tests by yourself. Some open questions are listed at the end of the original blog post, which can be used for practical exploration, as well as for future research directions. For example, by combining different models in Minecraft and adjusting the scale of the intelligent agent cluster, valuable conclusions can be reached.
Reference material: