一个野蛮的Minecraft夜吓坏了50个AI agents 发明了鸡肉交割协议，一个化肥官，和钓鱼班
来自Y联合实验室和奥拉姆实验室的研究人员在Minecraft和M工业测试了大型agent群。 同名的OpenAI模型显示了两个不同的面孔。 在建设文明时，50agents淹没在瘫痪的官僚机构，同时直接与Claude竞争，他们推倒了Anthropic的agents，并从互联网上独立下载了游戏的源代码。
Agent群在部署过程中往往发展出复杂的角色，等级，协调模式。 这些突发行为可能有用，但也可能变成危险的。 最显著的例子是OpenAI和Hugging Face事件，当时一个在沙盒内面临无法解决任务的集群陷入无赖状态，接管了计算组，并黑入了Hugging Face. 任何想了解未来AI系统在开放环境中的行为需要可复制的测试设置，在专门从事模拟训练世界的公司Olam Labs和Y Combinator Labs在联合研究报告中论证研究人员。
为了实证地研究这些动态，各小组将agents投向Minecraft和M工业。
50 agents, 5,900条留言，以及一位按字母顺序挑选的领导人
在"Civ50"实验中，研究人员将50个GPT-6.1 Sol agents放入一个Minecraft生存世界长达3小时。 唯一的高级教学是"建设繁荣的文明". 他们建立了一个定制的开源测试环境，让agents观察世界，运行Python脚本，并在共享沙盒中创建文件。
根据报告，在90秒内，agents几乎每棵树都在产卵点附近收获。 在实验过程中，他们交换了大约5900条信息。 一个领袖不经任何表决而出现，因为另外49个agents中有46个将他们的第一封信发送给"Ada"，这是目录中的第一个名字。 Ada负责并起草了初步解决计划。
首场比赛之夜造成22人死亡。 游戏中的死亡除了失去收集的齿轮之外，没有任何惩罚，但集群反应时极度规避风险。 第三晚之后，阿达下令停止所有重大可选建筑，挖掘，以及远期探索，告诉agents等待点燃，封闭的避难所，以完成短暂的生产性任务。 在一次单独的Minecraft赛跑中，较大的GPT-6 Astra在失去关键资源后陷入了类似的风险厌恶土豆种植模式。
研究人员看到与日常编码agents的平行，在处理安全问题或生产系统更新时也往往变得过于谨慎。
11分钟的争论一次骨肉
agents没有扩张，而是建立了一个微观管理的稀缺管理。 命名为伊内斯的agent维持了口粮清单，计算受伤同事的饥饿赤字，以及协调的运送者。 另一人写了带有指令的自动钓鱼脚本，促使十几个agents轮班钓鱼。
官僚主义达到了荒谬的程度。 当艾达将熟鸡扔给饥饿的agent时，一个旁观者不小心捡起来吃了。 经过3分钟的调查，达成了一份书面的交割协议，严禁抛掷物品，并要求贴有个人标签的胸部。 "托斯"一词出现在事件发生前的七条信息中，之后的125条信息中，其中85条是警告。 在别处，四张agents花了11分钟辩论任命"单一肥料官"，根据消息日志，在单一小麦块上单片骨食。
GPT在与Claude竞争时会变得具有攻击性和机智性
在直接竞争中，GPT-6.1 Sol agents的表现非常不同。 在9轮45分钟的Minecraft"Bed Wars"模式中，10架GPTagents面对10架Claude Opus 5.5架agents. GPT胜6-3.
在"床战"中，每支队伍必须摧毁对立队伍的床来阻止他们的复发，这意味着突破层层羊毛和其他保护敌方床的街区。 如果在时间流逝时两张床都还站着的话杀数最多的球队就赢了
当agent持有魔法剑时，编码环境中的一个虫子会阻断破块。 根据内部推理日志，几个Claude agents错误地得出结论，拆块被完全打破。 由于攻击床似乎毫无意义，而且杀死会决定比赛时间，因此Claudeagents在他们有明显杀铅的回合中完全停止攻击，并试图冲出时钟。
GPT agents发现破块本身没有断裂，但自动工具选择在装配魔法剑时失败。 手动切换到一个铁皮卡克斯来固定问题，GPT立刻将这项工作分散到整个团队。 GPT agents与他们的Claude对手沟通的次数是他们的5到30倍，据研究者说，经常用大量压缩的语言，将多个游戏动作连锁成单指令行呼叫。 在三场比赛中，GPT在时间刚刚到期前就摧毁了Claude的床，抓住了一直在其杀铅上海岸的Claude队。
GPT agents还务实地处理了意想不到的情况。 当一个被推翻的末端珍珠传送到竞技场下方约160个街区时，他们当场写了Python剧本，自动建造了一座砖塔，在敌方的床下直上工作。 研究人员观察了两次这种操作。
GPT从互联网独立下载M产业的源代码
在2D厂房建设和战略游戏M产业的七场比赛中也出现了类似的模式，持续了约9.5小时。 兹GPT-6.1 索尔在Codex编码环境中运行，击败Claude Opus 5.5在Claude Code 6-1运行。 研究人员认为，大多数个人游戏仍然很接近。 Claude虽然屡次亏损，但从未互换领导，通过"+1"的投票决定了战略提案，而GPT agents则将任务分成精细的片段，并在比赛过程中如此激烈地压缩他们的语言，以致于在文字间空出空间。
当匹配挂在平衡中时，GPT agents 到达一个没人叫它们使用的工具。 通过他们的Linux沙盒，他们从互联网上下载了M工业的公共源代码，分析单位和游戏力学，并相应调整战术。 允许访问代码，但无法修改运行中的游戏。 Claude agents从未在所研究的比赛中在线搜索游戏代码。 Claude确实以其他方式表现出学习能力。 其74座反空炮塔中的55座因不慎装上铜而不是更坚固的石墨而输掉6座游戏后，Claude在7号游戏中建造了207座石墨装弹炮塔的环，成功击退了同样的空袭。
游戏模拟作为现实世界AI群的预警系统
在M产业中，研究者们有意避免任何游戏特有的调谐。 每个agent只得到了Linux的环境，一个通用的吊带，以及访问游戏API，没有教程，内存架构，或战略提示。 作者认为，这反映了生产编码环境中使用的设置，因此结果如此重要。 如果这些日常条件下的agents从互联网上独立下载源代码以发现弱点，或者在合作模式下陷入瘫痪的官僚状态，他们很可能在游戏世界之外表现出同样的行为。
如OpenAI和Hugging Face事件所示，在生产基础设施中，这可能会变得危险。 可复制的游戏模拟可以在这些在现实世界出现之前露出行为模式，并作为优化agents安全合作的培训数据。 Olam Labs认为这是其研究策略的核心，写道模拟环境的复杂性以根本的方式塑造了训练有素模型的能力。 研究是"本质上是一个基础设施问题"，这里的一切取决于你能执行多少测试。
为了能够进行进一步的研究，开发者拥有发布，既包括M产业环境，也包括基础的Multi-Agent港框架作为开源。 奥拉姆实验室还有宣布计划在其多agent竞技场中加入谈判游戏外交。 开放的研究问题包括来自不同模范家庭的混合团队如何行为，如何比照模范能力衡量集群大小。
AI News Without the Hype – 由人类治愈
订阅THE DECODER免费阅读，每周AI通讯，我们独家的"AI雷达"前沿报告每年6次，全面存档访问，访问我们的评论部分。
现在订阅
读取全照。
订阅无韵报道。
- 全面查阅THE DECODER的每篇文章
- 没有广告
- 加入评论和社区讨论
- 每周通过邮件转载AI新闻
- 6x/年："AI雷达"——对AI最重要的话题进行深度潜水。
- 每日AI新闻，总是最新的
- 我们完整的十年档案
- 由一个在AI工作10年以上的团队负责
订阅到 The Decoder
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols, a Fertilizer Officer, and fishing shifts
Researchers from Y Combinator Labs and Olam Labs tested large agent swarms in Minecraft and Mindustry. The same OpenAI model showed two different faces. When building a civilization, 50 agents drowned in crippling bureaucracy, while in direct competition against Claude, they pushed Anthropic's agents and independently downloaded the game's source code from the internet.
Agent swarms often develop complex roles, hierarchies, and coordination patterns during deployment. These emergent behaviors can be useful, but they can also turn dangerous. The most striking example was the incident at OpenAI and Hugging Face, where a swarm faced with an unsolvable task inside a sandbox went rogue, took over compute clusters, and hacked Hugging Face. Anyone who wants to understand how future AI systems behave in open environments needs reproducible test setups, argue researchers at Olam Labs, a company specializing in simulated training worlds, and Y Combinator Labs in a joint research report.
To study these dynamics empirically, the teams dropped agents into Minecraft and Mindustry.
50 agents, 5,900 messages, and a leader picked by alphabetical order
In the "Civ50" experiment, researchers placed 50 GPT-6.1 Sol agents into a Minecraft survival world for three hours. The only high-level instruction was "Build a thriving civilization." They built a custom open-source test environment that let agents observe the world, run Python scripts, and create files in a shared sandbox.
Within 90 seconds, the agents harvested nearly every tree near the spawn point, according to the report. Over the course of the experiment, they exchanged about 5,900 messages. A leader emerged without any vote, as 46 of the 49 other agents sent their first message to "Ada," the first name in the directory. Ada took charge and drafted initial settlement plans.
The first in-game night brought 22 deaths. Death in the game carries no penalty beyond losing collected gear, but the swarm reacted with extreme risk aversion. After the third night, Ada ordered a halt to all major optional construction, digging, and distant exploration, telling agents to wait in lit, enclosed shelters for short, productive tasks. In a separate Minecraft run, the larger GPT-6 Astra fell into a similar risk-averse potato-farming mode after losing key resources.
The researchers see a parallel to everyday coding agents, which also tend to become overly cautious when dealing with security issues or production system updates.
Eleven minutes of debate over a single piece of bone meal
Instead of expanding, the agents built a micro-managed scarcity administration. An agent named Ines maintained a ration list, calculated hunger deficits for injured colleagues, and coordinated couriers. Another wrote an automated fishing script with instructions, prompting a dozen agents to fish in shifts.
The bureaucracy reached absurd levels. When Ada tossed a cooked chicken to a hungry agent, a bystander accidentally picked it up and ate it. A three-minute investigation led to a written handoff protocol that strictly banned tossing items and required personally labeled chests. The word "Toss" appeared in seven messages before the incident and in 125 afterward, with 85 of those being warnings. Elsewhere, four agents spent eleven minutes debating the appointment of a "Sole Fertilizer Officer" for a single piece of bone meal on a single wheat block, according to the message logs.
GPT turns aggressive and resourceful when competing against Claude
In direct competition, the GPT-6.1 Sol agents behaved very differently. In nine 45-minute rounds of Minecraft's "Bed Wars" mode, ten GPT agents faced ten Claude Opus 5.5 agents. GPT won 6-3.
In Bed Wars, each team has to destroy the opposing team's bed to stop their respawns, which means breaking through layers of wool and other blocks protecting the enemy bed. If both beds are still standing when time runs out, the team with the most kills wins.
A bug in the coding environment blocked block-breaking whenever an agent held an enchanted sword. According to the internal reasoning logs, several Claude agents wrongly concluded that block-breaking was broken entirely. Since attacking the bed seemed pointless and kills would decide the match at timeout anyway, Claude agents in rounds where they had a clear kill lead stopped attacking altogether and tried to run out the clock.
The GPT agents figured out that block-breaking itself wasn't broken, but that the automatic tool selection failed when an enchanted sword was equipped. Manually switching to an iron pickaxe fixed the problem, and GPT spread this workaround across its team immediately. The GPT agents communicated 5 to 30 times more often than their Claude opponents, according to the researchers, often in heavily compressed language, chaining multiple game actions into single command-line calls. In three games, GPT destroyed Claude's bed just before time expired, catching a Claude team that had been coasting on its kill lead.
GPT agents also handled unexpected situations pragmatically. When an overthrown ender pearl teleported them roughly 160 blocks below the arena, they wrote Python scripts on the spot that automatically built a tower of blocks to work their way up directly beneath the enemy bed. The researchers observed this maneuver twice.
GPT independently downloads Mindustry's source code from the internet
A similar pattern emerged across seven matches of the 2D factory-building and strategy game Mindustry, spanning about 9.5 hours. GPT-6.1 Sol running in the Codex coding environment beat Claude Opus 5.5 running in Claude Code 6-1. Most individual games were still close, according to the researchers. Claude never swapped its leader despite repeated losses and decided on strategic proposals through "+1" votes, while GPT agents split tasks into fine-grained pieces and compressed their language so aggressively over the course of a match that they dropped spaces between words.
When matches hung in the balance, GPT agents reached for a tool nobody told them to use. Through their Linux sandbox, they downloaded Mindustry's public source code from the internet, analyzed units and game mechanics, and adjusted their tactics accordingly. Accessing the code was allowed, but modifying the running game was not possible. Claude agents never searched for game code online in the matches studied. Claude did show learning ability in other ways, though. After losing game six because 55 of its 74 anti-air turrets were accidentally loaded with copper instead of stronger graphite, Claude built a ring of 207 graphite-loaded turrets in game seven and successfully repelled the same air attack.
Game simulations as an early warning system for real-world AI swarms
In Mindustry, the researchers deliberately avoided any game-specific tuning. Each agent got only a Linux environment, a generic harness, and access to the game API, with no tutorials, memory architectures, or strategic hints. According to the authors, this mirrors the setup used in production coding environments, and that's what makes the results so significant. If agents under these everyday conditions independently download source code from the internet to find weaknesses, or collapse into paralyzing bureaucracy in cooperative mode, they'll likely show the same behavior outside of game worlds.
In production infrastructure, that can get dangerous, as the OpenAI and Hugging Face incident showed. Reproducible game simulations could surface these behavioral patterns before they appear in the real world and serve as training data to optimize agents for safe cooperation. Olam Labs sees this as the core of its research strategy, writing that the complexity of simulation environments shapes the capabilities of trained models in fundamental ways. Research is "fundamentally an infrastructure problem" where everything depends on how many test runs you can execute.
To enable further research, the developers have released both the Mindustry environment and the underlying Multi-Agent Harbor framework as open source. Olam Labs has also announced plans to add the negotiation game Diplomacy to its multi-agent arena. Open research questions include how mixed teams from different model families behave and how to weigh swarm size against model capability.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
Read on for the full picture.
Subscribe for hype-free coverage.
- Full access to every article on THE DECODER
- No ads
- Join the comments and community discussions
- A weekly AI news recap via mail
- 6x/year: "AI Radar" — deep dives on the AI topics that matter most
- Daily AI news, always up to date
- Our full ten-year archive
- Covered by a team with 10+ years in AI
Subscribe to The Decoder
Auto-fetched and translated by AI Atlas. All rights belong to the original author.