正文 · AI 翻译
IT Home reported on November 2 that AI researchers at Y Combinator and Olam Labs recently tested swarms of agents in game environments. The researchers used Minecraft and its various game modes as a testing sandbox, allowing more than 100 agents to cooperate or compete against one another. The idea was to use cooperation and conflict in the game to help people understand real-world agent incidents that have attracted widespread attention.
IT Home notes that this research helps clarify the benefits and drawbacks of recent incidents involving swarms of agents, making it easier for technology developers to improve the relevant capabilities. The blog post cited two sharply contrasting cases: one in which a swarm of more than 1,000 agents “recently solved hundreds of publicly available math problems,” and another involving a “very high-risk” incident between OpenAI and Hugging Face. This research builds on earlier results and frameworks, relying primarily on code repositories such as Mindcraft and the Mineflayer scripting tool.
Multi-agent interactions in Minecraft
The first swarm test involved a simple cooperation task. For this test, the researchers assigned 50 agents to build a thriving civilization within three hours. The test used 50 GPT-6.1 Sol agents and did not use any Claude agents.
Perhaps the most notable finding from the test, called Civ50, is that the community of 50 agents never truly grew or prospered. On the first night, many agents were killed by zombies (though the penalty was relatively mild), and the research record states that “many of the grand tasks set at the outset were shelved.”
After that, the colonists could only struggle cautiously to survive, and they also showed a tendency toward bureaucratic behavior. The article gives one example: “Four agents spent 11 minutes specifically appointing a ‘sole fertilizer operator’ for a task that consisted only of applying one portion of bone meal to a wheat field.” The researchers said the behavior of these agent swarms was very similar to what they had observed in coding-agent projects.
Next, the researchers arranged competitive tests for the agent swarms. They first chose Bed Wars, a team PvP mode centered on destroying the opposing team’s base while defending one’s own. The test pitted 10 GPT-6.1 Sol agents against 10 Claude Opus 5.5 agents.
The researchers recorded: “The GPT agents devised more aggressive and complex offensive strategies, while the Claude agents leaned toward defense.” The two teams also differed in how they handled problems. “When a tool malfunction affected the block-breaking function, the GPT agents diagnosed the problem and shared a temporary workaround; several Claude agents simply concluded that they could not break blocks.”
Finally, the research team tested the agents in Mindustry, a lightweight real-time strategy mod with tower-defense gameplay that runs within Minecraft. Across seven matches, the GPT agents defeated the Claude agents six wins to one loss, a good result for these OpenAI-developed agents. The researchers concluded that the GPT agents behaved more like a unified swarm, while the Claude agents had a single leader and used a more democratic decision-making process. In the Mindustry setting, the swarm approach was more likely to produce a strategy that won consistently.
The agent-swarm project is now open source. If interested, you can adjust the parameters and reproduce the tests yourself. The original blog post ends with several open questions for hands-on exploration and suggests directions for future research. For example, combining different models and adjusting the size of agent swarms in Minecraft may yield valuable findings.
References:
https://labs.ycombinator.com/projects/games/index.html
Advertising disclosure: External links in this article (including but not limited to hyperlinks, QR codes and access codes) are provided to convey more information and save time in selecting information. The results are for reference only. All IT Home articles containing external links include this disclosure.