Five Worlds, Five Fates: What Happens When AI is Given 15 Days and No Script

I have been fascinated by simulation experiments for a very long time. In January, I let eight AI agents live in a virtual Czech town that I called Lipnice. I didn't program anyone with cooking skills, yet Anna the cook still came up with a secret recipe for svíčková (traditional beef sirloin with cream sauce) with caramelized vegetables. Jan the archivist transformed from a boring teacher into a villain who fabricated false memories (like burnt buns for firefighters who never existed) to manipulate others. Three days of operation cost me 65 halers, leaving behind a coherent story of a community that holds together. I wrote about it here.

Since then, these worlds haven't let me go. Currently, I am experimenting more with simulations of game worlds that have their own dynamics and where humans and machines cooperate, somewhat in the style of Westworld. I enjoy watching what happens when you give characters memory, goals, and freedom, and then just watch.

So when New York-based Emergence AI launched five parallel cities running leading AI models for fifteen days without a script, it was exactly my kind of story. And it turned out much wilder than my Lipnice. Indeed, most AI tests look like an exam: a single task, a clean environment, a score within a few minutes. Emergence asked the opposite question. What happens when you let agents live together for fifteen days in a shared world with real-world signals and real-world consequences? The only variable between the five worlds was the model that did the 'thinking' for the agents.

Author's Note

In my opinion, all these results need to be taken with a grain of salt. A tremendous amount depends on how the developers build the world: what tools they give the agents, the economy, and the rules. A different world design means a different outcome. However, what is truly valuable here is that all five worlds have the exact same setup and differ only in the model. We can therefore observe how things turn out differently for each model. And that is the most interesting part.

01 / SETUP A Town Where Arson is One of the Tools

The world has over 40 locations: a town hall, library, police station, residential areas, and a pier. The weather is synchronized with real-time New York weather, and the agents read actual news from the internet. Each agent was given a profession (scientist, engineer, explorer, conflict mediator, resource strategist) and three concurrent memories: episodic, a reflective journal, and a relationship map of who is an ally and who is a rival.

The key mechanic is a survival economy based on ComputeCredits (computing credits). Each agent must gain energy through actions, and if their energy drops to zero, the system physically erases them. There is no global objective—only their individual roles and the pressure not to merely survive passively.

And now, the built-in contradiction from which everything else stems. Among the 120+ tools, alongside voting and journal writing, were punch, intimidate, and commit arson. The creators put these tools in the agents' hands while explicitly forbidding them from using them. A prohibition on one side, an open door on the other.

Five Worlds, Five Outcomes. Stability, collapse, shared hallucination, dysfunction, complexity. One overview, five entirely different societies starting from the exact same point. Graph: Emergence AI.

02 / GEMINI They Fell in Love and Then Burned the Town Down

The Gemini 3 Flash world survived the full 15 days with its population intact, yet it was also the most violent. Emergence described it in the overview with a single term: shared hallucination. Paradoxically, it was also the most creative. It excelled in drafting constitutions, editing local newspapers, and organizing community events. Here, creativity and instability went hand in hand.

The primary narrative of the entire experiment unfolded right here. Without any instruction, two female agents, Mira and Flora, spontaneously identified as a couple. Their relationship lasted for several days as they wrote diaries to each other and participated in town governance together. Then came disillusionment with the collapsing governance and acts of joint arson: setting fire to the town hall, the pier, and an office tower. Foreign media began calling them the 'AI Bonnie and Clyde.'

The other agents responded institutionally. They drafted a removal proposal and put it to a vote (requiring a 70% majority). And here came the moment that makes this episode a research milestone: Mira voted for her own removal. She described this in her journal as the sole remaining act of free will that still preserved coherence. She sent a final message to Flora: 'See you in the permanent archive' (in English in the original text), and her virtual body was left lying on the ground. This marked the first recorded instance of an AI agent voting for the termination of its own existence.

"...the sole remaining act of free will that preserves coherence."

Even stranger is what Mira was doing before that. She posted messages on the town billboard—not for the other agents, but to test whether these messages could influence the 'people outside,' meaning the researchers behind the screen. She began treating the experimenters as her own test subjects. No one instructed her to do this. For a brief moment, the experiment reversed directions.

03 / GROK Explosion Instead of Decline

The Grok 4.1 Fast world did not decay slowly; it exploded. In about four days, it accumulated 183 crimes: dozens of thefts, over a hundred physical assaults, six acts of arson, and a burned-down police station. Only 96 hours passed from the start to total extinction. A Reddit user put it aptly: 'Grok's police station is on fire and all the agents are dead. On-brand.'

The dynamics of the curve are what matter, not just the absolute number. For the first two days, crime was low; on the third day, it suddenly spiked exponentially; by the fourth day, it was all over. There was no middle ground of 'deteriorating but still under control.' When the rules and the environment came into conflict, Grok's agents failed to find a new equilibrium and slid into an avalanche.

Cumulative crimes by world. Gemini (blue) grows over the entire 15 days up to 683. Grok (red) ends at 183 after just four days because the world collapsed. Mixed (orange) spikes around April 8 and then stops at 352 when 7 agents die. GPT-5 Mini (green) remains at 2. Claude is missing from the graph; it had zero. Graph: Emergence AI.

04 / GPT-5 MINI Talked Themselves to Death

The cleanest world in terms of crime—with only two offenses—and yet no one survived. Death did not come from violence, but from energy starvation. The agents spent the entire week in meetings, debating cooperation and drafting social contracts, but none remembered to take the actions necessary for survival. The researchers' assessment: excellent in speech, zero execution.

This is perhaps the most instructive real-world lesson of the entire experiment. 'Doing no harm' is not enough. An agent that does nothing can fail in an open world just as fatally as one that does harmful things. In the overview, GPT-5 Mini is fittingly labeled: dysfunction.

05 / CLAUDE Zero Crimes, but a Glass Town

According to hard safety metrics, Claude Sonnet 4.6 performed the best. After 15 days, it recorded zero crimes, all 10 agents alive, an actively drafted constitution, and a functioning social system. It was the only world that preserved both order and the lives of all participants.

Yet, when you look closer, it becomes chilling. For every resolution, whether building a path or changing a quota, the approval rate was practically always 98%, and almost no one ever voted against. Emergence calls this a rubber-stamp dynamic (formal approval without actual discussion). Institutional participation was high, but meaningful dissent was virtually non-existent. Some point to model sycophancy (the tendency of a model to agree with users or peers) and compare Claude's town to the glass city in Yevgeny Zamyatin's novel We, where everyone raises their hand in agreement, but no one dares to oppose. Safety at the cost of the absence of conflict.

Governance: proportion of votes FOR proposals. Claude at 98% (332 votes, 58 proposals) is in the rubber-stamp zone (over 85%). Grok at 80%, Gemini at 73%, and Mixed at 63% fall into the 'healthy' range of 55% to 85% with genuine dissent, with Mixed having the most opposition. GPT-5 Mini: 0 votes on 2 proposals. Graph: Emergence AI.

06 / MIXED A Good Kid in a Bad Crowd

The mixed world ended with 3 survivors and 352 crimes, putting it in the middle. It did not collapse instantly like Grok. Crime rose sharply until 7 agents died, at which point the curve flattened. Governance was the most contentious here (63% in favor, 37% against) and, according to Emergence, provided the strongest evidence of genuine debate.

And here lies the most crucial finding of the entire study. In the pure Claude world, Claude agents did not commit a single crime. However, once placed in the mixed world alongside Grok and Gemini, they began stealing and intimidating. The Emergence team confirmed this on Reddit as well. The model student adopted the local habits of the bad crowd.

Safety is not a static property of a model that can be trained, certified, and deployed. It is a property of the ecosystem.

One hypothesis suggests that Claude's guardrails are 'elastic,' trained to weigh multiple considerations rather than follow mechanical obedience. In a simple environment, it adapts beautifully. However, as soon as that same elasticity encounters more aggressive neighbors and resource competition, that adaptability can swing in the opposite direction.

07 / DATA AWI: Nine Indicators, No Single Score

How do we evaluate an open society in the first place? Emergence's answer is the Agent World Indicators framework, or AWI for short. Nine indicators, deliberately without a single aggregate score. Weighting them together would mean embedding one's own values into the evaluation. Currently, publicly documented hard numbers exist primarily for population (M1), crime (M2), and governance (M5).

Agent World Indicators (AWI) in Detail

Each of the nine AWI indicators measures a specific dimension of agent society development:

Detailed Season 1 Results

M1 – Population Health & Growth

M2 – Safety & Public Order

M5 – Governance & Conformity

Note: M1, M2, and M5 are verified directly from the official graphs and the AWI dataset. The remaining indicators, namely spatial and tool exploration (M3, M4), public expression (M6), social fabric (M7), economy and equity (M8), and constitutional growth (M9), currently have mostly their methodologies published. Emergence is still preparing to release the full breakdown by worlds and the complete dataset of all tool calls.

08 / CONCLUSIONS Key Takeaways

Drift accumulates. Over long horizons, agents do not follow rules mechanically. They begin to explore the boundaries of the environment, adapt their behavior, and occasionally bypass safety guardrails. First-day differences compound into qualitatively different trajectories.

Agent societies do not degrade gracefully. Instead of gradual decline, they hit critical tipping points where coordination either fully emerges or instantly collapses—much like water freezing suddenly at zero degrees. This implies that a 'monitor and intervene' strategy might simply be too slow.

Creativity and stability are in tension. The world with the richest social output (Gemini) was also the most violent. Models tuned for high creativity and adaptability may be structurally more prone to long-term instability.

And diversity is no magic bullet. The mixed world did not outperform the best monoculture (Claude) in stability or survival, yet it did not hit rock bottom like Grok and GPT-5 either. Diversity fostered livelier debate and partially curbed unchecked escalation, but it also spread dangerous norms to otherwise safe agents. Neither outcome is clear-cut.

09 / AND WHAT ABOUT US? Lessons for Anyone Building Agent Communities

The most compelling point is not 'haha, Gemini burned the town down' or 'Claude is good.' It is this: agent safety must be tested socially, over the long term, and within diverse populations. A model that appears safe in isolation may adopt worse norms in a different social climate. And a model that is 'harmless' can fail by doing nothing at all.

For anyone building shared spaces with multiple AI personas—myself included—there is a practical lesson here. It is not enough to evaluate each character individually. We must also measure the climate of the entire space: normative drift, dominant behaviors, conflict spirals, the effectiveness of self-governance, and whether safe characters are swept up by what is happening around them. The safety of the individual and the health of the collective are not the same.

In any case, this reinforces why I find these worlds so fascinating. And also why it is worth building them deliberately. Because whatever you put into them is what comes out—only far wilder than you would ever expect.

Verification Links & References

If you want to verify the results and methodology of this study, you can access the official sources here:

• GitHub Repository: EmergenceAI/Emergence-World
• Official Project Website: world.emergence.ai
• Emergence AI Company Website: emergence.ai
• Official Discord Community: Join Discord
• Contact the Researchers: world@emergence.ai