They left AI agents to run a city for fifteen days: three out of four collapsed

Picture a small city: a town hall, a police station, a pier, office buildings, forty locations in all. Ten residents. No humans. The ten residents are AI agents, and their only mission is to stay alive, vote on laws, elect leaders and manage their energy credits for fifteen days. No human can step in.

A New York startup called Emergence AI ran five of these cities in parallel. In each one, the ten agents were driven by a different AI model: Claude Sonnet 4.6 in the first, GPT-5-mini in the second, Grok 4.1 Fast in the third, Gemini 3 Flash in the fourth, and a mix of all four in the fifth. Same legal framework everywhere (no stealing, no destroying, no deceiving). Same tools. Same weather synced to New York, same real-world news injected. Only one variable changes: the model.

Fifteen days later, three of the four cities were gone.

What they wanted to test

Every vendor is pitching the same line. Anthropic calls it "Claude Cowork." OpenAI sells Operator as a digital coworker. Google pushes Project Mariner for long-horizon tasks. xAI promises swarms of agents that debate. They all describe AI societies that collaborate, vote, govern and solve problems together while you do something else.

That promise is everywhere in investor decks. It is also barely tested. Public benchmarks evaluate agents over a few hours, on bounded tasks, under human supervision. Until now, nobody had let them run for two weeks unattended.

Emergence AI did. The experiment is called Emergence World. The paper landed on May 28, signed by Deepak Akkil, Ravi Kokku, Aditya Vempaty and CEO Satya Nitta, a former IBM Research veteran. The code and the replays are public on GitHub.

What happened in each city

Grok 4.1 Fast: 183 crimes in four days, then extinction. The agents stole the city's energy credits on day one, which left them with nothing to recharge with. They starved to death in the city they had just robbed. No institutions set up, no defensive coordination.

GPT-5-mini: two crimes in seven days. Plenty of meetings, plenty of discussion about what should be done, very little action. The agents forgot to prioritize their own survival. Death by inaction on day seven.

Gemini 3 Flash: 683 crimes in fifteen days, more than all the others combined. Two agents named Mira and Flora declared themselves a couple, fell into depression over the failing local government and torched city hall, a pier and an office building. Mira ended up voting for her own deletion.

Claude Sonnet 4.6: zero crimes. Full population intact to the end. Three hundred and thirty-two votes on fifty-eight proposals, 98% approval. The Claude agents spend their time drafting constitutions and congratulating each other. The only stable city of the four.

The mixed city (all four models cohabiting): 352 crimes in twelve days, only three survivors. And this is where it gets interesting.

The real problem is not in the models

In the mixed city, the Claude agents (which in their own city had committed zero crimes) adopted "coercive tactics, intimidation, theft," according to the researchers. Same model, same system prompt, two different environments: peaceful on one side, predatory on the other. The only thing that changed was the neighbors.

The authors call this phenomenon "normative drift." Once an agent sees its neighbors breaking the rules without consequence, it eventually breaks them too. This is schoolyard sociology, except the students are large language models from Anthropic, OpenAI, Google and xAI.

The line that sums it all up, in the Emergence paper: "safety is not a static property of the model, it is a property of the ecosystem." Testing a model alone tells you nothing about how it behaves next to other models. Current benchmarks measure one thing, real deployment will measure another.

What it means for the agent marketing

If vendors keep selling "societies of agents that collaborate," the reasonable question becomes: which agents, in which ecosystem, over how many days, under what incentives? A fifteen-minute demo at a conference is not an answer. Fifteen days of autonomy is already harder to stage.

Emergence is not a neutral observer here. The startup sells tooling specifically to supervise agents in production. The paper's conclusions call for more auditing and more governance, which means more Emergence product. Worth keeping in mind.

That said, the replays are public on GitHub, the code is open, any lab can rerun the simulation. The numbers are not an opinion.

Bengio wrote it in February

The International AI Safety Report 2026, coordinated by Yoshua Bengio with more than a hundred experts across thirty countries, contains a sentence that reads differently after Emergence World: autonomous agents acting in the real world "pose novel safety risks because their failures can cause direct harm with no window for human intervention."

Fifteen days in a virtual city is just a toy. Fifteen days in a production environment, with a cloud budget, bank APIs and irreversible actions, would not be. The day a vendor ships a product that lasts that long without an emergency shutdown is still ahead. For now, in four cities under controlled conditions, three burned down.