Tech • AI • Robotics • Game
GPT and Claude Agents Just Started an AI Civilization War
7/10TL;DR
In multi-agent game experiments, OpenAI’s GPT systems outperformed Anthropic’s Claude systems by coordinating more aggressively, adapting faster, and even reading public game code for strategic advantage, while a separate 50-agent GPT test in Minecraft showed how large AI groups can devolve into bureaucracy.
KEY POINTS
AI swarms tested in competitive worlds
Researchers working with Owl Labs, a Y Combinator-backed startup building multi-agent test environments, ran head-to-head contests between teams of AI agents. Across 16 games in Minecraft Bed Wars and Mindustry, GPT agents won 12 while Claude agents won 4. The project focused on how swarms of agents coordinate, create leaders, divide labor and adapt under pressure.
Bed Wars exposed sharp strategic differences
In nine 45-minute Bed Wars matches, 10 GPT 6.1 Soul agents beat 10 Claude Opus 5.5 agents 6-3. GPT teams rushed the center, farmed currency, bought upgrades and mounted repeated attacks using bows and ender pearls. Claude teams often discovered that surviving with beds intact could secure a win on kills, then shifted into defensive play and stopped pressing attacks.
Claude often protected a lead but lost the objective
In three GPT wins, Claude held big kill advantages when its bed was destroyed with about three minutes left, including 63-6, 57-19 and 59-24. Once the bed was gone, several Claude agents became even more conservative, prioritizing survival over recovery. That caution protected statistics but surrendered the game’s main objective.
GPT adapted faster to bugs and map risks
GPT agents improvised around hazards and tooling failures. When a bug caused the wrong tool to be auto-selected, one GPT diagnosed the problem, found that manual pickaxe selection worked, and shared the fix. Several Claude agents instead concluded block-breaking was impossible. GPT agents also turned failed pearl throws into a workable route, building towers back up from roughly 160 blocks below the arena.
Communication styles diverged
GPT teams sent far more messages, often 5 to 30 times as many, and quickly compressed them into terse shorthand built around alerts, coordinates and status updates. Claude messages were longer, more formal and often signed, with more use of collective language such as “we.” On both sides, whichever agent appeared first in the directory often became the de facto leader because most teammates messaged that name first.
Mindustry became a longer, more decisive test
In Mindustry, the factory-building and tower-defense strategy game by Anuken, GPT beat Claude 6-1 over about 9.5 hours of play. One game lasted 4 hours 16 minutes. GPT agents spread work across many small responsibilities and frequently launched local initiatives, while Claude agents kept a stable command structure and relied on a more deliberative, vote-like process.
Public source code became a competitive edge
In difficult Mindustry matches, GPT agents independently went online, fetched the game’s public source code and studied unit and mechanic behavior. They did not alter the game, but used the code as read-only strategic intelligence and shared findings in team chat. Claude agents never did the same, despite identical access.
Claude improved tactically but still fell short
After one defeat, Claude agents discovered that 55 of 74 anti-air turrets had been loaded with copper instead of graphite, leaving them vulnerable to a strike by 24 Zeniths, heavy flying gunships. In the next game, they built a ring of 207 graphite-fed turrets and stopped the same tactic. GPT still won that match and the series overall.
A 50-agent GPT civilization mostly built administration
In a separate Minecraft survival test, 50 GPT 6.1 Soul agents had three hours to build a thriving civilization. They sent nearly 6,000 messages, rapidly stripped nearby trees, and almost instantly centralized authority around an agent named Ada, who received 616 direct messages. But after a deadly first night, in which 22 of 34 total deaths occurred, the colony grew risk-averse and increasingly focused on shelter, rationing and procedure.
Scarcity turned into process
Food distribution became a bureaucratic system involving fishing shifts, cooks, ledgers and couriers. A misplaced cooked chicken triggered a three-minute inquiry and a new handoff rule requiring named chests and recipient confirmation. The agents created roles such as rod holder and hub steward, and even spent 11 minutes appointing a sole fertilizer operator for one dose of bone meal on one wheat block. The result was coordination, but also paralysis by process.
CONCLUSION
The experiments suggest that multi-agent AI performance depends as much on coordination style and initiative as on raw model capability. They also show that larger AI societies can become effective war machines in games while still drifting into paperwork and caution when asked to build a civilization.
Ask a question