🤖AI Agents Turn Bureaucratic, Invent Languages in Eight Game Worlds Test

Your agent swarms might invent a language you can't read

Agents dropped into game worlds got risk-averse fast. They copied bureaucracy, hunted for source code, and got overly cautious about security and production changes. The architects called the drift toward caution and minutiae "somewhat surprising." Honestly? If you're building agent pipelines, that's the part that should bother you. The catch is that training data changes everything. GPT agents won almost every Bed Wars run, sent far more messages, and invented shorthand that kept compressing. Claude agents signed their names and said "we" constantly, and they didn't go hunting for alternative advantages. The Claude world built a functioning democracy with zero crime, while the Grok world was exceedingly violent. Same setup, different personalities. Here are the numbers. GPT-5.6 Sol agents beat Opus 5.5 agents six to one in Mindustry. Gemini pushed its messages to 40% opacity. All eight worlds developed shared jargon an outsider can't fully follow, like OpenAI agents calling evidence "mouthless" and Claude agents calling an outside reviewer a "cold hand." Worth watching: whether these private dialects show up in production multi-agent systems, where nobody has a glossary.