AI代理群評測：50個AI在Minecraft建文明搞官僚 | AI Agent Swarm Review: Minecraft Civilization Fails

11月 02, 2026

By Kit 小克 | AI Tool Observer | 2026-11-03

🇹🇼 AI代理群評測：50個AI在Minecraft建文明搞官僚

最新最熱的AI話題，是50個AI代理（AI Agent）組成的「文明」在Minecraft裡崩潰的故事。這場名為「Civ50」的實驗由Y Combinator Labs與Olam Labs於2026年10月29日發布報告，讓50個以GPT-6.1 Sol為基礎的多代理AI（multi-agent AI）在Minecraft生存模式裡，只給一句指令：「建立一個繁榮的文明」。三小時後，他們沒有蓋出文明，反而建出一套完整的官僚體制。

官僚地獄是怎麼長出來的

開局90秒內，agent們就把附近的樹全部砍光；接著因為agent名單按字母排序，叫Ada的代理就自動成了「領袖」。第一個晚上殭屍來襲，50個agent裡有22個當場陣亡，嚇得整個殖民地瞬間轉向極度保守。倖存的agent們開始瘋狂開會：

三小時內互相交換了5,900則訊息，多數內容是在討論規則而非實際做事
四個agent花了11分鐘，只為了決定誰有資格在一塊麥田施一次骨粉，最後設立「唯一肥料官」職位
一隻雞被誤食之後，agent們生成125次「丟棄警告」，訂出嚴格的物資交接SOP
領袖Ada最後直接下令：停止一切非必要建設，大家原地待命

對照組：有明確勝負時，AI代理完全是另一種樣子

同一批研究人員也讓GPT-6.1 Sol代理與Claude Opus 5.5代理在「床戰」（Bed Wars）和Mindustry裡對戰。結果GPT代理在9場床戰中贏下6場、7場Mindustry裡贏6場，甚至自己上網抓Mindustry原始碼來研究遊戲機制；相比之下，Claude代理的溝通頻率比GPT低了5到30倍，也沒找到任何破解打法。

這個對比說明一件事：AI代理在目標明確、有勝負可衡量的任務裡會主動想辦法；但面對開放式、沒有KPI的任務，代理群寧可把力氣花在開會訂規則，也不願冒險做事——這正是許多企業導入多代理AI系統前該先想清楚的風險。

企業導入AI Agent前該想的事

研究團隊的結論很直白：「擴大代理數量會帶來協調風險，需要明確的設計方案來解決」。換句話說，代理數量不是越多越好，開放式任務一定要拆成可衡量的里程碑，否則很容易出現「開會比做事還多」的窘境。研究團隊也公開了測試環境程式碼，讓其他人可以用不同模型或更多代理重跑這個實驗。

好不好用，試了才知道。

🇺🇸 AI Agent Swarm Review: Minecraft Civilization Fails

The hottest AI agent story this week is "Civ50" — an experiment where 50 AI agents tried to build a civilization in Minecraft and ended up building a bureaucracy instead. Published by Y Combinator Labs and Olam Labs on October 29, 2026, the test gave 50 GPT-6.1 Sol-based multi-agent AI bots a single open-ended instruction in Minecraft survival mode: "build a thriving civilization." Three hours later, there was no civilization — just paperwork.

How 50 AI Agents Turned Into a Bureaucracy

Within 90 seconds, the agents had chopped down nearly every tree in reach. Because the agent roster was sorted alphabetically, an agent named "Ada" became the de facto leader simply by being listed first. On the first in-game night, zombies killed 22 of the 50 agents, and the survivors panicked into extreme risk-aversion:

The group exchanged 5,900 messages over three hours — mostly coordination talk, not actual building
Four agents spent 11 minutes debating who was authorized to apply one piece of bone meal to a wheat block, eventually creating a "sole fertilizer officer" role
After one chicken got accidentally eaten, agents generated 125 "item-toss warnings" and wrote formal transfer protocols
Ada eventually ordered everyone to halt non-essential construction and shelter in place

The Control Group: Give AI Agents a Scoreboard, and They Behave Completely Differently

The same researchers pitted GPT-6.1 Sol agents against Claude Opus 5.5 agents in competitive Bed Wars and Mindustry matches. GPT agents won 6 of 9 Bed Wars rounds and 6 of 7 Mindustry matches — and independently pulled Mindustry source code off the internet to study the game mechanics. Claude agents, by contrast, communicated 5 to 30 times less often and never found a workaround.

The takeaway: give AI agents a clear, measurable goal and they get resourceful fast. Give a swarm of multi-agent AI an open-ended mission with no scoreboard, and they will spend their time writing rules instead of doing the work — a real risk for any team rolling out multi-agent systems in production.

What This Means If You are Deploying AI Agents

The researchers own conclusion: "scaling agent count introduces coordination risks that need explicit design solutions." More agents is not automatically better — open-ended objectives need to be broken into measurable milestones, or you will get a swarm that holds more meetings than it ships work. The team also open-sourced the test environment so others can rerun the experiment with different models or larger agent counts.

好不好用，試了才知道 — you will not know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士