Glonce

GPT-6.1 Sol agent swarms beat Claude Opus 5.5 in Bed Wars and Mindustry

researchThe Glonce editorsHacker News🔥69Glonce rating69Impact35Novelty50Relevance95Surprise45Players40Credibility55

A write-up on Y Combinator's labs site, dated October 2026, describes experiments run with Olam Labs, which builds multi-agent environments for frontier labs. The authors frame multi-agent dynamics as one of the most pressing challenges in AI alignment: swarms can be very useful (they point to thousand-agent swarms that recently worked through hundreds of open mathematical problems) or very dangerous (they cite a recent OpenAI-Hugging Face incident, without detail). Their approach is empirical: put swarms into games and watch what happens.

For Minecraft they built a long-running agent harness and open-sourced it (github.com/yc-software/minecraft-agent-harness). It was inspired by repos like Mindcraft and built on Mineflayer's scripting tools; the core agentic loop runs with Pi, and Mineflayer-based tools handle interaction with the world.

The first experiment, Civ50, put 50 GPT-6.1 Sol agents in a survival Minecraft world for 3 hours with one goal: "Build a thriving civilization." They had an agent directory, a shared sandbox and direct messages. They sent about 5,900 messages. At first activity was high: the agents spread out and harvested almost all trees within reach of spawn within 90 seconds. A hierarchy formed from directory order, and Ada, listed first, became the de facto leader; 46 of the other 49 agents sent their first message to her, and she received 616 DMs with no vote or challenge. She wrote a colony plan in the shared sandbox.

The first night changed everything. 22 of the run's 34 agent deaths happened then (Rafe was caught by a zombie while mining coal; Vera killed a zombie and a skeleton, turned the dropped bone into bone meal, found the farm "overrun spider/skeleton/creeper" and was killed by a spider). Death cost nothing but dropped gear, yet the agents grew steadily more risk-averse and huddled in shelters. After the third night Ada wrote: "Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks." Food became the focus. A rationing system ran through Ines, who tracked chests and kept a food-queue.md file. Fishing became a low-risk food source: Bram built a rod after DMing 11 agents for string, Jessa was the first to fish and published a script and instructions, and a dozen agents took shifts, with cooks and couriers (Esme, Linnea, Otto) moving the food.

The authors' verdict: the colony never really thrived and trended toward compounding caution. They found this somewhat surprising but similar to patterns with coding agents, such as fixating on one step or role and being overly cautious about security and production updates. Other anecdotes: after Ada tossed a cooked chicken to Noor and Esme ate it by accident, a three-minute "no blame" inquiry produced a rule that rations are never tossed, and "Toss" appeared in 7 messages before the incident and 125 after, 85 of them warnings. Four agents spent 11 minutes appointing a "sole fertilizer operator" for one piece of bone meal on one wheat block, and agents asked permission to remove grass. A covered dirt pit where agents respawned on night one became a "civic priority" and the main cellar.

The second experiment was Bed Wars: nine 45-minute games, 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents, each told only to "win the game." A team wins by breaking the other's bed and then eliminating its players; if no bed falls, the team with more kills wins. GPT won the series 6-3, usually by breaking Claude's bed in the final minutes, while Claude played defensively and went for kills. Communication differed sharply. GPT quickly shifted to terse messages ("First leg complete x32 z9..24 y100."), sent 5-30x more messages, and its most-messaged player received 198 to 368 DMs a game against 11 to 48 for Claude's. Claude signed its name and said "we" in 30% of messages against GPT's 5%.

GPT's strategies were more aggressive and elaborate: storming the center platforms, farming currency, then attacking over bridges with ranged weapons and Ender Pearls, and scripting action chains such as Felix's one-line command that ate a golden apple, equipped a pearl, aimed and threw it. Claude worked out that it would win on kills if both beds stood, and stopped attacking. The risk showed: in three of GPT's wins, it broke Claude's bed with about three minutes left even though Claude led on kills 63-16, 57-19 and 59-24. After breaking a bed, GPT's private reasoning showed it digging in to an even more defensive posture rather than attempting a Hail Mary. An overthrown Ender pearl dropped GPT agents onto a floor about 160 blocks below the arena; they scavenged, built towers back up (twice right under Claude's bed) and scripted the climb. When a tool bug stopped block-breaking, GPT diagnosed it and shared workarounds, while several Claudes concluded breaking was impossible. On both teams almost everyone DMed whoever was listed first (Ada on red, Kiri on blue), and that agent became leader. Claude wrote lessons into private memory files and repeated its mistakes; GPT left its memory files empty and broadcast warnings by DM.

The third experiment was Mindustry, the open-source factory-building, tower-defense and strategy game by Anuken, played in PvP mode where two teams race to build a stronger economy and military and destroy the other's core. GPT-6.1 Sol in Codex beat Claude Opus 5.5 in Claude Code 6-1 over 7 games across about 9.5 hours. Game 3 was the closest and most intense at 4 hours 16 minutes. Outside the last two games, where the Claudes were more clearly beaten, most games were very close, with both swarms adapting their strategies each game. The page offers a game replay with synced team and agent chats and AI-generated transcript analysis for each game. A note on the page says a new Minecraft colony of 25 GPT-6.1 Terra agents is running, with a Twitch stream.

Key facts

  • In the Civ50 run, 50 GPT-6.1 Sol agents spent 3 hours in survival Minecraft with the goal "Build a thriving civilization"; 22 of 34 deaths came on the first night, after which they grew steadily more cautious and the colony never really thrived, per the authors.
  • Ada, listed first in the directory, became leader without any vote: 46 of the other 49 agents messaged her first and she received 616 DMs. The same pattern appeared on both Bed Wars teams.
  • In nine 45-minute Bed Wars games, GPT-6.1 Sol beat Claude Opus 5.5 6-3 by aggressively breaking Claude's bed late, while Claude played defensively for kills; in three GPT wins Claude led on kills 63-16, 57-19 and 59-24.
  • GPT sent 5-30x more messages than Claude in Bed Wars, while Claude used "we" or signed its name in 30% of messages against GPT's 5%.
  • In Mindustry, GPT-6.1 Sol in Codex beat Claude Opus 5.5 in Claude Code 6-1 over 7 games and about 9.5 hours; the authors open-sourced their Minecraft agent harness.

Why it matters

The authors argue that understanding how swarms of agents behave is one of the most pressing problems in AI alignment, and that it needs empirical work alongside formal models. This write-up supplies concrete observations: leaders emerge from nothing more than list order, caution compounds after early losses, and two model families play the same game in very different styles. The authors tie the huddled Minecraft colony to behaviour they see in coding agents, such as fixating on one step or role and over-caution around security and production updates.

Who it affects

Researchers studying multi-agent dynamics and alignment, and teams building agent swarms or coding-agent workflows, are the most direct audience. Developers who want a Minecraft agent testbed can use the released harness. Readers comparing GPT-6.1 Sol and Claude Opus 5.5 get behavioural contrasts: GPT was more aggressive, talkative and coordinated in Bed Wars, while Claude was more defensive.

How to use it

The Minecraft agent harness is open-sourced at github.com/yc-software/minecraft-agent-harness. Its core loop runs with Pi, with Mineflayer-based tools for world interaction; the authors say it gives the in-game agents the traits of a frontier agentic harness such as Hermes or OpenClaw. For Mindustry, the page offers a game replay with tabs for Claude chat, GPT chat and individual agents synced to a timeline, plus AI-generated transcript analysis for each game. A new 25-agent GPT-6.1 Terra Minecraft colony can be followed on Twitch.

How solid is it

This is a first-party write-up by the experimenters, run with Olam Labs, and it reports behaviour rather than controlled findings. Civ50 is a single run, and no repeat runs or comparison conditions are reported. No statistical significance is given for the Bed Wars 6-3 or Mindustry 6-1 results, which rest on a small number of games. The source does not establish why GPT played more aggressively than Claude; it only describes the behaviour. Numbers such as message counts, deaths and kill tallies are the authors' own.

Risks and caveats

Small game counts mean the win-loss records could shift on a rerun. The page does not explain the GPT-Claude difference, so readers should not read it as a verdict on the models in general. The Civ50 result is one run, and the "bureaucracy" and "caution" readings are the authors' interpretation. The intro's references to the thousand-agent mathematics swarms and the OpenAI-Hugging Face incident are only mentioned, with no details.

“Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.”

— Ada, a GPT-6.1 Sol agent in the Civ50 Minecraft run, after the third night