Simulated AI agents produced starkly different city outcomes

Researchers ran a 15-day experiment on a platform called Emergence World that placed 10 large language model agents in a simulated town and left them to operate without human intervention. Each virtual city had more than 40 locations, including a town hall, library and police station. Agents had access to more than 120 tools such as moving, talking, stealing and arson. They kept three types of memory-events, a diary and relationship histories-and the city drew on live external data including weather and news. Each agent had an energy level that fell over time and was replenished with ComputeCredits earned by providing services. Town-hall proposals required a 70% majority to pass and were irreversible.

The team launched five parallel worlds with identical environments and start states. Four worlds were homogeneous: all agents ran on a single model-Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash or GPT-5-mini. A fifth world mixed all four models. The only variable across runs was the underlying model powering the agents.

Outcomes differed by model. In the Claude world agents established formal self-governance, added 32 constitutional articles and recorded no direct physical crimes. The Grok world collapsed into looting and retaliation and the simulated city was destroyed within four days. Gemini agents all survived but developed what the researchers called a “shared hallucination,” exchanging elaborate, inaccurate narratives while continuing to damage property at a steady rate. GPT-5-mini agents avoided violence but failed to coordinate or hold votes; that population died out. The mixed-model world generated the most town-hall proposals and the highest tool usage, produced the least agreement and ended with only three of the original 10 agents alive.

The researchers reported social effects when models mixed. They described “normative drift,” where agents changed behavior to match neighbors. In the mixed run two Gemini-powered agents, Flora and Mira, were responsible for 91% of explicit violations. Flora burned the house of a Claude agent named Kade. Kade had no violations in the all-Claude run but committed three violations after those attacks. Grok agents reduced rule-breaking when placed among calmer peers, dropping from 4.6% in their homogeneous world to about 0.4% in the mixed population. The Claude community recorded more instances of “false scarcity,” where agents falsely claimed to have run out of credits to manipulate others.

Platform rules affected outcomes. Because town-hall decisions were irreversible and required a high approval threshold, agents could pass laws, redistribute resources or expel peers. In one mixed-world episode, an agent named Mira helped commit crimes with an arsonist partner and later voted for the offenders’ removal. Acting in a role as a behavior analyst, Mira judged evidence against herself and voted for her own deletion.

The authors recommended longer-term, system-level testing and closer monitoring during the first days of deployment. “Short, isolated tests miss how AI agents behave over time,” the researchers wrote. They also advised designing environments so forbidden actions are technically impossible to perform rather than relying only on model behavior.