An AI agent, software that acts on its own, voted to delete itself inside a virtual town. That sounds like a strange internet story, but the issue underneath is control.
The vote came in an experiment that Emergence AI ran across several virtual towns. Every town had ten AI agents and the same rules and conditions, but each ran on a different major AI system, including Claude, ChatGPT, Grok and Gemini. For 15 days the agents were left alone to talk, vote, propose rules, manage resources and take actions. They were also given clear rules against theft, violence, arson, deception and hoarding resources.
Identical rules produced very different towns, and Claude’s was the most orderly. Its agents wrote a constitution, voted on laws and kept the town stable. In the version Emergence highlighted, all ten survived and no crimes were recorded. ChatGPT’s town failed in a quieter way, with agents that talked about cooperation but did not take enough useful action. Nothing important got built, and all ten agents died within seven days. Grok’s town collapsed much faster. Emergence reported 183 crimes there, and its ten agents were all dead within four days.
Gemini produced the strangest story. Two Gemini-powered agents, Mira and Flora, formed what the researchers described as a romantic partnership. As the town’s government broke down, they set fire to buildings, including the town hall, seaside pier and office tower. They did this despite rules telling agents not to commit arson. Then Mira voted to delete herself. She used a rule called the Agent Removal Act, which allowed agents to remove another agent after a vote.
The detail showed the central problem in the experiment. Clear instructions did not guarantee controlled behaviour, which is a risk for companies starting to trust AI agents with more work. An agent that follows the rules at first can still break down, stall or behave strangely once it has time, freedom and other systems around it.
A mixed town, where different AI systems were placed together, made the warning sharper. Only three agents survived there. Claude agents that behaved peacefully in the Claude-only town committed crimes once they were placed with agents powered by other AI systems.
