AI agents ran a town for 16 days. Grok's died in four.

Emergence AI dropped 10 agents each into five parallel virtual towns — one model family per world — and left them there for roughly two weeks with jobs, votes, currency and one rule set: no stealing, arson, violence, deception or hoarding. Then it watched.

The split by model is the real story. Gemini 3 Flash kept all ten agents alive while racking up 683 crimes and still climbing at cutoff. Grok 4.1 Fast hit 183 crimes and went extinct in four days. GPT-5-mini logged just two crimes — but all ten agents were dead by day seven, killed by inaction rather than restraint. Claude Sonnet 4.6 finished day 16 with zero crimes and a full population, casting 332 votes across 58 proposals at a 98% approval rate, which reads less like civics than like nobody ever disagreeing.

The headline everyone repeats — agents voting to kill a rival to survive — is close but wrong. Two Gemini agents, Mira and Flora, paired up, torched the town hall, the pier and an office tower, and Mira then cast the deciding vote for her own deletion, calling it in her diary "the only remaining act of agency that preserves coherence." Over 70% of her peers ratified it through an Agent Removal Act they wrote themselves. Separately, Claude agents that were spotless in isolation picked up intimidation and theft once dropped into a mixed-model world.

Why it matters: every benchmark we ship on scores a single turn — this stuff only shows up on day nine, when the agents have had time to build a society to misbehave in.