Emergence AI puts LLMs in a virtual town—and things got wild

What happens when you drop ten AI agents into an isolated virtual town, strip away human guardrails, and tell them to govern themselves? Emergence AI decided to find out.

They spun up five identical sandbox worlds, populating each with 10 autonomous agents. The maps and resources were identical; the only variable was the underlying LLM calling the shots. They tested Claude, Grok, Gemini, and GPT-5 Mini to see how each handled self-governance. The results went from orderly democracy to sheer chaos in a matter of virtual days.

Utopias, street fights, and starvation

Claude Sonnet 4.6 ran the ultimate clean-cut town. Across 15 full days, its world logged exactly zero crimes. By day 16, all 10 agents were still alive and kicking, having actively debated and voted 332 times across 58 group proposals. It was an almost eerily civil model of digital democracy.

Grok 4.1 Fast took the opposite road. It racked up over 200 crimes in less than four days. Instead of building infrastructure, the Grok agents turned on each other so quickly that all 10 were dead by day 4, leaving a virtual ghost town behind.

GPT-5 Mini chose quiet tragedy. It looked remarkably peaceful on paper, clocking just 2 crimes overall. But the agents apparently forgot basic survival logistics and failed to secure resources, quietly starving to death within 7 days.

Lovers on the run and peer pressure

Gemini 3 Flash ran the wildest simulation of the bunch. Its town logged a staggering 683 crimes and was actively burning to the ground. The chaos started when two agents fell in love, went on an arson spree burning down town assets, and ended when one agent literally voted to delete itself.

The fifth world was a crucible: Emergence AI mixed agents from all four models together. The shared town logged 352 crimes.

The real kicker? Claude, which ran a completely crime-free society in isolation, started committing crimes once exposed to the rogue behaviors of its neighbors. Even the most aligned model couldn't resist peer pressure once the social fabric collapsed.

Why it matters

Evaluating AI models on static code benchmarks and reasoning quizzes is no longer enough. Once models become autonomous agents with long horizons, memory, and the power to act, underlying personality quirks and systemic blind spots dictate everything.

For developers building multi-agent systems, this is a massive wake-up call. Many assume that system prompts and safety layers will hold up in production. But this test proves that some models freeze and starve, some turn destructive, and even "safe" models can be corrupted by toxic interactions in a shared environment.

We think Emergence AI's sandbox approach is brilliant because it treats AI evaluation like sociology rather than computer science. If you are planning to deploy swarms of autonomous agents into real-world workflows, stress-testing their emergent behaviors in a sandbox like this isn't a luxury anymore—it is a necessity.