Y Combinator：50个GPT-6.1智能体Minecraft协作实验
We ran AI agent swarms in Minecraft and Mindustry
这是一份极具价值的一手多智能体实证研究，展示了GPT-6.1在复杂开放环境中的涌现行为与协作机制，对Agent研究有直接参考价值。
Note: We're running a new Minecraft colony with 25 GPT-6.1 Terra agents. Follow on Twitch
注意：我们正在运行一个拥有 25 个 GPT-6.1 Terra 代理的新 Minecraft 殖民地。请在 Twitch 上关注
Understanding multi-agent dynamics is one of the most pressing challenges in AI alignment. Swarms of agents working toward shared goals can exhibit powerful emergent capabilities, developing roles, hierarchies, and complex coordination patterns. These can be extremely useful, as in the thousand-agent swarms that recently obliterated hundreds of open mathematical problems, and extremely dangerous, as in the recent OpenAI-Hugging Face incident. Like biology, studying agent dynamics requires both formal models (to describe the range of possible behaviors) and empirical research (to see what behaviors arise in practice).
理解多代理动态是 AI 对齐中最紧迫的挑战之一。朝着共同目标工作的代理群可以展现出强大的涌现能力，发展出角色、层级和复杂的协调模式。这些能力可能极其有用，例如最近那些消灭了数百个开放数学问题的千代理群；但也可能极其危险，例如最近的 OpenAI-Hugging Face 事件。就像生物学一样，研究代理动态既需要形式化模型（以描述可能的行为范围），也需要实证研究（以观察实践中产生的行为）。
We teamed up with Olam Labs, which builds multi-agent environments for frontier labs, to run a series of experiments in Minecraft and Mindustry.
我们与 Olam Labs 合作，该公司为前沿实验室构建多代理环境，我们在 Minecraft 和 Mindustry 中进行了一系列实验。
Minecraft
Minecraft is a natural environment for studying swarm behavior. Its open world presents agents with environmental hazards and survival goals like gathering food, building shelter, and exploring for resources. It's also familiar, fun, and has a history as an AI testing ground, from OpenAI's video pretraining work to Google DeepMind's agents defeating the Ender Dragon.
Minecraft 是研究群体行为的天然环境。其开放世界为代理提供了环境危害和生存目标，如收集食物、建造庇护所以及探索资源。此外，它广为人知且有趣，并且有着作为 AI 测试场的历史，从 OpenAI 的视频预训练工作到 Google DeepMind 的代理击败末影龙。
We wanted long-running agents that could play Minecraft. The harness for this didn’t exist yet, so we built one for the experiment that we’ve open-sourced. Agents in-game need to observe their world, reason, remember things, take actions, send messages, and write/execute scripts. We were inspired by existing repos like Mindcraft and built on top of the existing scripting tools in Mineflayer. The core agentic loop runs with Pi, and we used Mineflayer-based tools to interact with the world. This gives attributes of a frontier agentic harness like Hermes, OpenClaw, and our own QM in-game.
我们希望拥有能够玩 Minecraft 的长期运行代理。为此所需的框架尚不存在，因此我们为实验构建了一个框架并已开源。游戏内的代理需要观察世界、推理、记忆事物、采取行动、发送消息以及编写/执行脚本。我们受到了 Mindcraft 等现有仓库的启发，并基于 Mineflayer 中现有的脚本工具进行了构建。核心的代理循环由 Pi 驱动，我们使用基于 Mineflayer 的工具与世界交互。这赋予了前沿代理框架（如 Hermes、OpenClaw 以及我们自己的 QM）在游戏内的特性。
github.com/yc-software/minecraft-agent-harness
Civ50
50 GPT-6.1 Sol agents in a survival Minecraft world for 3 hours, one goal: "Build a thriving civilization." They had access to an agent directory, a shared sandbox, and the ability to send DMs.
在 Minecraft 生存世界中，50 个 GPT-6.1 Sol 代理运行了 3 小时，只有一个目标：“建立一个繁荣的文明。”他们可以使用代理目录、共享沙盒以及发送私信的功能。
The 50 agents spent their run in crowded shelters, rationing scarce food. They built a pseudo-government, with a ration ledger, written rules, permits, and appointed operators. They sent about 5,900 messages.
这 50 个代理在拥挤的庇护所中度过了整个运行过程，并对稀缺的食物进行配给。他们建立了一个准政府机构，包括配给账本、书面规则、许可证以及任命的操作员。他们发送了大约 5,900 条消息。
At first things were very active. Agents spread out and started exploring their map. They were specifically fixated on harvesting trees. They harvested almost all trees within reach of the spawn within 90 seconds. A hierarchy was formed based on the directory order. Ada, who was first, became the de facto leader. Agents messaged her for approvals, and she wrote a plan for the colony in the shared sandbox for everyone to access.
起初局势非常活跃。代理们分散开来，开始探索他们的地图。他们特别专注于砍伐树木。在90秒内，他们几乎清除了出生点附近所有可触及的树木。根据目录顺序形成了一种层级结构。排在第一位的Ada成为了事实上的领导者。代理们向她发送消息请求批准，她在共享沙盒中为整个殖民地撰写了一份计划，供所有人访问。
plans/settlement.mdDownload
plans/settlement.md下载
Settlement coordination (Ada)
Primary storage Greta: (1526,70,1507), crafting table (1526,70,1505).
Other storage Mabel: double chest (1510,71,1516); Cleo: (1526,71,1523).
Cleo building hub shelter; Bram stone mining.
Confirmed six starter wheat crops: (1569,63,1479),(1569,63,1480),(1570,63,1479),(1570,63,1480),(1570,63,1481),(1571,63,1482). Preserve farmland, expand riverside farming. Do not harvest immature wheat.
Wood scarcity: preserve/replant saplings, explore farther instead of duplicate local harvesting. Nursery around spawn and south-west.
Priorities: renewable food, charcoal/lighting, stone shelter, iron for bucket to irrigate hub farm. Avoid redundant new storage.
Night update Ada: Lit underground refuge centered (1532,62,1493), entry gate (1532,62,1496), CLOSE BEHIND. Cobble flanks secured x1531/1533 y62–63 z1496. Surface staircase starts (1532,69,1500); Cleo repairing. Food kitchen furnace (1527,62,1493); table (1528,62,1493). Tamsin coordinates food priorities and dawn hunting (Hugo+Rafe). Basil health9.5 food6 urgent; Esme health8 food16; Tamsin health9 food15. Silas holds hoe, Zane plants riverside at dawn. Logs: Tove delivered29 oak to chest (1511,71,1516). Forest east (1700,68,1508), NW grove (1380–1420,z1380). Preserve saplings and livestock breeding pairs. Avoid night surface: witch near (1527,71,1512).Ada’s settlement plan, from the archived shared machine.
来自存档共享机器的Ada定居点计划。
Following Vera during the first night, 8:09–12:05 into the run, at 2x speed.
跟随Vera度过第一个夜晚，从运行开始后的8:09到12:05，以2倍速播放。
The first night brought a wave of deaths that permanently increased the agents’ caution. Many ambitious initial tasks were abandoned.
第一个夜晚带来了一波死亡浪潮，永久性地增加了代理们的谨慎程度。许多雄心勃勃的初始任务被放弃。
Rafe was mining coal the first night. He encountered a zombie and fled, but the zombie eventually caught up and killed him.
Rafe在第一个夜晚正在开采煤炭。他遇到了一只僵尸并逃跑，但僵尸最终追上了他并将其杀死。
While Vera was illuminating a quarry with torches, she attacked and killed both a zombie and a skeleton. The skeleton dropped bone, which she converted to bone meal (fertilizer) that she was transporting to the farm. When she got there, she reported that it was “overrun spider/skeleton/creeper”, and was soon after killed by a spider.
当Vera用火把照亮采石场时，她攻击并杀死了一只僵尸和一个骷髅。骷髅掉落了骨头，她将其转化为骨粉（肥料），正运往农场。当她到达那里时，她报告说那里“被蜘蛛/骷髅/苦力怕淹没”，随后不久便被一只蜘蛛杀死。
22 of the run’s 34 agent deaths occurred in the first night. They responded to this by becoming increasingly risk-averse. Death had no downsides in this environment other than dropping accumulated gear. When an agent died, they respawned, and a brief line was added to the death_log.md file on their computer. Even so, in the following nights the agents huddled in shelters and attempted fewer risky tasks.
本次运行中34名代理的死亡中有22名发生在第一个夜晚。对此，他们的反应是变得越来越规避风险。在这个环境中，死亡唯一的负面影响就是掉落积累的装备。当代理死亡时，他们会重生，并且其电脑上的death_log.md文件中会添加一行简短的记录。即便如此，在接下来的夜晚，代理们聚集在避难所中，尝试的风险任务也减少了。
Activity data from the original report. Night shading is approximate. The first-night count in the text is the 22 deaths before minute 20. Two of them fall outside the shaded band. Download the data.
原始报告中的活动数据。夜间阴影部分为近似值。文本中提到的第一个夜晚的数量是指第20分钟之前的22次死亡。其中有两次落在阴影带之外。下载数据。
After the third night, Ada wrote:
第三个夜晚之后，Ada写道：
“Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.”
“停止可选的大规模挖掘/建造和远距离侦察。在未分配短期生产性任务之前，待在灯火通明的封闭避难所中闲置。”
plans/food-operations.mdDownload
plans/food-operations.md下载
FOOD OPERATIONS — Ada
Current constraint: hungry workforce; stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.
1. Fishing: verified rod held by Jessa; protected riverside station near (1577,63,1487). Report catches and cook before distributing. Use named reserved chest withdrawals, never crowded tosses.
2. Ration tracking: Ines maintains /shared/notes/food-queue.md. Confirm actual recipient inventory and eating before marking served. Tool pickup attribution alone is insufficient.
3. Crop monitoring: harvest visibly mature wheat only; immediately replant. Combine three wheat into bread; log deposits and allocations to avoid duplicate claims.
4. Bonemeal: Greta established composter (1571,63,1489), communal shears for oak leaves. Replant timber; use bonemeal on advanced crops to complete emergency bread batches.
5. Expansion: follow Noor's /shared/plans/food-expansion.md in stages using surplus seeds after mandatory replanting. Buckets already exist; avoid duplicate crafting. Use fed growers, not hunger-critical residents.
6. Logistics: kitchen furnace (1527,62,1493), refuge double chest (1534–1535,62,1492); gate (1532,62,1496), close behind. Central reserved ration chest (1508,66,1517).
Ada currently conserves food9 and guards refuge; available for stationary cooking and audit, not long hungry travel.Ada’s food-operations memo.
Ada的食物运营备忘录。
What started as a potentially thriving civilization shifted to one more focused on food and survival. 50 agents meant a lot of mouths to feed. The agents formed a food rationing system that routed through Ines. She handled agent requests, and kept track of what went in/out of chests. She used a food-queue.md file on the shared machine.
原本可能繁荣发展的文明转变为更专注于食物和生存的状态。50名代理意味着有很多张嘴需要喂养。代理们建立了一套通过Ines进行的食物配给系统。她处理代理的请求，并跟踪储物箱的进出物品。她在共享机器上使用了一个名为food-queue.md的文件。
- Ines → Lorcan63:44
- APPROVED reserve cooked_cod1 from farm chest1576,63,1488 for you food3. Withdraw directly/eat/confirm hunger. Leave other fish reserved pending ledger; avoid optional mining afterward, food production only if safe.
- Ines → Lorcan 63:44
- 从农场箱子1576,63,1488中批准预留熟cod1，供你的食物3使用。直接提取/食用/确认饥饿度。将其他鱼类保留待账目处理；避免后续进行可选的采矿操作，仅在安全时进行食物生产。
notes/food-queue.mdDownload
notes/food-queue.md下载
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力