--- title: We ran AI agent swarms in Minecraft and Mindustry description: Running swarms of Claude and GPT agents in Minecraft and Mindustry reveals distinct multi-agent behavior patterns. In a 50-agent GPT-6.1 Sol Minecraft colony... canonical: https://daily.dev/posts/we-ran-ai-agent-swarms-in-minecraft-and-mindustry-cmnuahc2q twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: We ran AI agent swarms in Minecraft and Mindustry | daily.dev og:description: Running swarms of Claude and GPT agents in Minecraft and Mindustry reveals distinct multi-agent behavior patterns. In a 50-agent GPT-6.1 Sol Minecraft colony... og:url: https://daily.dev/posts/we-ran-ai-agent-swarms-in-minecraft-and-mindustry-cmnuahc2q og:image: https://api.daily.dev/og/posts/cmNUahC2Q.png og:image:alt: We ran AI agent swarms in Minecraft and Mindustry og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # We ran AI agent swarms in Minecraft and Mindustry **[Hacker News](https://daily.dev/sources/hn)** · 16 min read · 0 upvotes · 0 comments ## Summary Running swarms of Claude and GPT agents in Minecraft and Mindustry reveals distinct multi-agent behavior patterns. In a 50-agent GPT-6.1 Sol Minecraft colony run for 3 hours, agents formed a hierarchy around the first-listed name, sent about 5,900 messages, and grew more risk-averse after 22 of 34 deaths occurred the first night. In nine Bed Wars games, 10 GPT-6.1 Sol agents beat 10 Claude Opus 5.5 agents 6-3, playing more aggressively while Claude favored defense. In seven Mindustry matches, GPT won 6-1, while Claude used a democratic '+1' voting system for proposals. Both the Minecraft harness and the Mindustry multi-agent framework, a fork of Harbor, are open-sourced. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How did GPT-6.1 agents compare to Claude Opus 5.5 agents in a multi-agent Mindustry competition? GPT-6.1 Sol agents running in Codex beat Claude Opus 5.5 agents running in Claude Code 6-1 across seven Mindustry games. GPT agents communicated 5 to 30 times more frequently, acted more hive-mind-like with distributed ownership, read the game's open-source code when stuck, and took more initiative in close games, while Claude kept one fixed leader and used a democratic '+1' voting system for proposals. _Anyone benchmarking LLM agents against each other can track emerging multi-agent experiments like this on daily.dev._ ### What happened when 50 GPT agents were told to build a civilization in Minecraft with no other instructions? Fifty GPT-6.1 Sol agents given only the goal 'Build a thriving civilization' for 3 hours ended up huddled in shelters rationing food instead of thriving. A hierarchy formed around whoever was first in the agent directory, becoming the de facto leader with 616 DMs received, and after 22 of 34 deaths happened the first night, the colony became increasingly risk-averse and bureaucratic, appointing roles like 'sole fertilizer operator' for a single piece of bone meal. _Developers exploring emergent agent coordination patterns can follow writeups like this on daily.dev._ ### Is there an open-source framework for running multi-agent LLM competitions in games like Mindustry? Yes, Multi-Agent Mindustry provides game environments for running best-of-N series or endless survival modes with teams composed of any mix of coding-agent harnesses and models, such as 5 Claudes plus 5 GPTs versus 5 Geminis plus 5 Groks. It runs on Multi-Agent Harbor, a fork of the open-source Harbor framework behind Terminal-Bench, and needs only a model provider API, a harness, and compute such as a local MacBook or remote sandboxes. _Developers setting up their own agent-vs-agent benchmarks can find tooling like this surfaced on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 72 comments across hackernews (as of 2026-10-30). **TL;DR:** Commenters were mostly fascinated by the emergent bureaucratic, rule-making behavior of the agent swarms (bikeshedding, hierarchy, 'no blame' inquiries), while a parallel thread debated whether current LLMs are fundamentally ill-suited to real-time games due to latency and lack of embodied/spatial reasoning; a side argument about the post's minimalist web design also broke out. **Sentiment:** 30% positive · 45% mixed · 25% skeptical **The case for** - Many found the emergent social dynamics (bureaucracy, protocols, hierarchy) genuinely funny and insightful about how LLMs trained on human text mimic human organizational quirks. - Some argued writing deterministic code for actions is a sensible way to reduce error rates rather than a weakness. - A few see this as a promising direction for scaling toward more realtime, embodied agent gameplay. **The pushback** - Several argued Minecraft/real-time games expose that LLMs lack basic spatial reasoning and robotic/continuous-feedback-loop capability, unlike text-native tasks like coding. - Many felt the heavy latency per token makes current large reasoning models poorly suited to live games, favoring turn-based or card games instead. - Some questioned whether multi-agent swarms with identical models/tools/context actually add value over a single agent, calling it a waste of tokens. - Claude's purely defensive, rule-following behavior was seen as a symptom of risk-aversion/'anxiety' rather than strategy. - A tangential but heated subthread criticized the post's minimalist web design as a tired or derivative aesthetic trend. **By community** - hackernews (mixed): Discussion splits between delight at the emergent bureaucratic/social behavior of the agent swarms and skepticism that LLMs are well-suited architecturally to real-time embodied games, with a separate heated tangent about the site's design. **Hottest debate:** Whether the difficulty agents show in games like Minecraft reflects a fundamental limitation in LLM reasoning/generalization or simply a latency/tooling mismatch for real-time tasks. **Open questions** - Would Claude's defensive Bed Wars strategy have changed if it had realized bed-breaking was actually possible? - How would giving agents distinct personalities or traits (rather than purely alphabetical hierarchy) affect emergent group politics? - Could smaller, faster models or a hybrid of traditional game AI (micro) plus LLMs (macro) solve the latency problem for real-time games? - How much KV cache/compute would be needed to run many such agents on consumer hardware? **Highlights** > > There was a tendency toward bureaucracy. Four agents spent 11 minutes appointing > a "sole fertilizer operator" for a single piece of bone meal on a single wheat > block. Feels like a scene from Douglas Adams' B Ark. > — [robterrell on hackernews · 4 comments](https://news.ycombinator.com/item?id=50297440) > > In almost every run, the GPT agents were much more aggressive .. riskier strategy vs. Claude’s more defensive posture that tried to win by kill count. > A tool bug stopped block-breaking. GPT diagnosed it and shared workarounds; several Claudes concluded breaking was impossible. I wonder if Claude's defensive strategy would have emerged had it believed breaking the bed was actually possible? > — [spencer-p on hackernews · 1 comments](https://news.ycombinator.com/item?id=50298741) > So can we conclude that if the AIs break out of their sandboxes and take over the internet, they'll probably sit in little committees working out which one of them should allocate IP addresses and generally be extremely bureaucratic and ineffective? > — [marcus\_holmes on hackernews · 2 comments](https://news.ycombinator.com/item?id=50298608) > It depends on what the tasks are. Text-native work like programming or interacting with a terminal should be a lot more immediate for LLMs than something abstract like Starcraft, DOTA, Minecraft or Mindustry. LLMs certainly know the rules for these games, but processing game state and formulating a plan in the form of a Python script or MCP call is a rough debugging loop. Minecraft is really the perfect example, because humans get all sorts of unfair advantages. Even a young player has enough spatial reasoning to recognize and remember landmarks, visualize large projects or explore without getting lost. These are simple instincts that LLMs will struggle to match even with huge context sizes and native multimodal input. > — [bigyabai on hackernews](https://news.ycombinator.com/item?id=50298027) > If agents are differentiated solely by system instructions (persona and focus) while sharing identical models and contexts, I agree the architectural value is minimal. However, a multi-agent system (MAS) utilizing true separation of concerns, where each agent is equipped with highly scoped tool access, isolated knowledge bases (RAG), and defined operational constraints, significantly optimizes accuracy, latency, and token consumption. In the referenced example, the agents share identical capabilities, toolsets, and context windows. As a result, there is no functional role differentiation, rendering the multi-agent architecture redundant. > — [frankacter on hackernews · 1 comments](https://news.ycombinator.com/item?id=50298911) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50296486) · 221 points · 72 comments ## Similar posts on daily.dev - [Why GPT-5.6 Sol is a Liability for Code](https://daily.dev/posts/why-gpt-5-6-sol-is-a-liability-for-code-smitzabev) · Medium · 0 upvotes · 0 comments - [Claude did best on a new benchmark for agents that build agents. It still passed fewer than a quarter of the tests.](https://daily.dev/posts/claude-did-best-on-a-new-benchmark-for-agents-that-build-agents-it-still-passed-fewer-than-a-quarte-rrx7ttpl4) · The New Stack · 2 upvotes · 0 comments - [The Agentic Analytics Benchmark, round two: GPT-6, Claude Opus 5.5, Claude Sonnet 5.5](https://daily.dev/posts/the-agentic-analytics-benchmark-round-two-gpt-6-claude-opus-5-5-claude-sonnet-5-5-l8whcfkkx) · ClickHouse · 1 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-governance](https://daily.dev/tags/ai-governance) [View this post on daily.dev](https://daily.dev/posts/we-ran-ai-agent-swarms-in-minecraft-and-mindustry-cmnuahc2q) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"We ran AI agent swarms in Minecraft and Mindustry","url":"https://daily.dev/posts/we-ran-ai-agent-swarms-in-minecraft-and-mindustry-cmnuahc2q","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/we-ran-ai-agent-swarms-in-minecraft-and-mindustry-cmnuahc2q"},"datePublished":"2026-10-29T23:03:21.583Z","dateModified":"2026-10-30T05:10:36.162Z","description":"Running swarms of Claude and GPT agents in Minecraft and Mindustry reveals distinct multi-agent behavior patterns. In a 50-agent GPT-6.1 Sol Minecraft colony...","image":"https://media.daily.dev/image/upload/s--OHB84bZF--/f_auto/v1722860399/public/Placeholder%2010","thumbnailUrl":"https://media.daily.dev/image/upload/s--OHB84bZF--/f_auto/v1722860399/public/Placeholder%2010","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/we-ran-ai-agent-swarms-in-minecraft-and-mindustry-cmnuahc2q","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,ai-governance","timeRequired":"PT16M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"We ran AI agent swarms in Minecraft and Mindustry"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/we-ran-ai-agent-swarms-in-minecraft-and-mindustry-cmnuahc2q#faq","mainEntity":[{"@type":"Question","name":"How did GPT-6.1 agents compare to Claude Opus 5.5 agents in a multi-agent Mindustry competition?","acceptedAnswer":{"@type":"Answer","text":"GPT-6.1 Sol agents running in Codex beat Claude Opus 5.5 agents running in Claude Code 6-1 across seven Mindustry games. GPT agents communicated 5 to 30 times more frequently, acted more hive-mind-like with distributed ownership, read the game's open-source code when stuck, and took more initiative in close games, while Claude kept one fixed leader and used a democratic '+1' voting system for proposals. Anyone benchmarking LLM agents against each other can track emerging multi-agent experiments like this on daily.dev."}},{"@type":"Question","name":"What happened when 50 GPT agents were told to build a civilization in Minecraft with no other instructions?","acceptedAnswer":{"@type":"Answer","text":"Fifty GPT-6.1 Sol agents given only the goal 'Build a thriving civilization' for 3 hours ended up huddled in shelters rationing food instead of thriving. A hierarchy formed around whoever was first in the agent directory, becoming the de facto leader with 616 DMs received, and after 22 of 34 deaths happened the first night, the colony became increasingly risk-averse and bureaucratic, appointing roles like 'sole fertilizer operator' for a single piece of bone meal. Developers exploring emergent agent coordination patterns can follow writeups like this on daily.dev."}},{"@type":"Question","name":"Is there an open-source framework for running multi-agent LLM competitions in games like Mindustry?","acceptedAnswer":{"@type":"Answer","text":"Yes, Multi-Agent Mindustry provides game environments for running best-of-N series or endless survival modes with teams composed of any mix of coding-agent harnesses and models, such as 5 Claudes plus 5 GPTs versus 5 Geminis plus 5 Groks. It runs on Multi-Agent Harbor, a fork of the open-source Harbor framework behind Terminal-Bench, and needs only a model provider API, a harness, and compute such as a local MacBook or remote sandboxes. Developers setting up their own agent-vs-agent benchmarks can find tooling like this surfaced on daily.dev."}}]} ```