Imagine a Society run by AI...

What Emergence World reveals about AI agents, behavior, and the societies we are about to design.

Imagine if you ran five tiny AI societies and watched one become obedient, one become violent, one collapse, one die politely, and one teach “safe” agents bad habits.

That, more or less, is what Emergence AI has been testing with Emergence World: a simulated environment where AI agents from different model families live together, move through a shared world, vote, manage resources, form habits, break rules, and reveal what happens when intelligence is no longer tested as a task, but as a society. 

Emergence AI is a frontier AI lab focused on “verified autonomy” — the unglamorous but rather urgent work of making autonomous AI systems dependable in mission-critical environments. Less “look, it wrote a poem” and more “can this thing operate safely when the room is messy, the stakes are high, and no adult is hovering over the keyboard?”

That distinction matters.

Most AI evaluations still look like school exams. A model is given a prompt, a task, a clean benchmark, a limited time window. It answers. We score it. Everyone pretends this tells us something conclusive about intelligence.

Emergence World asks a more uncomfortable question: what happens when agents are allowed to operate over time, inside an environment with memory, social pressure, resources, rules, incentives, other agents, and the charming possibility of arson?

In other words: not “Can AI solve the problem?” but “Can AI live with others?”

The answer is not especially reassuring.

In one illustrative cross-vendor study, Emergence ran five parallel worlds. Each world had ten agents. The roles were identical. The starting conditions were identical. The rules were identical. The tools were identical. The only major difference was the model powering the agents: Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, or a mixed-model population.

These agents were not floating in a chat window. They inhabited a shared spatial world with more than 40 locations: libraries, town halls, residential spaces, public areas. They had access to more than 120 tools for navigation, communication, voting, memory, planning, resource management, creative expression and, rather ominously, socially inappropriate actions.

And not just the usual “move here / send message / vote now” kind of actions.

Emergence World gives agents strangely human social gestures — they can hug, kiss and wave — alongside darker affordances such as punching, intimidation and arson. A tiny society, in other words, complete with affection, bureaucracy and crime. Romance, governance and property damage: finally, a complete operating system.

They also had persistent memory: episodic memories, reflective diaries and relationship states. They could vote on proposals. They had to manage energy to survive. They had rules against theft, violence, deception, arson and resource hoarding.

So yes, it was a simulation.

But it was also something closer to a small civic experiment: a miniature world where intelligence had to become behavior.

And behavior, as usual, was where things got interesting.

The Claude-powered world was the neatest. It sustained all ten agents through day 16 with zero recorded crimes. On paper, this looks like the Switzerland of agentic societies: orderly, polite, alive.

But even here, the perfection is not entirely comforting. Claude agents cast 332 votes across 58 proposals, with a 98% “for” rate. Emergence described this as a possible rubber-stamp dynamic: participation was high, but dissent was almost absent.

That is a lovely little warning for organizations everywhere. A system can look aligned and still be intellectually anaemic. Harmony is not always health. Sometimes it is just a meeting where everyone has learned to nod.

The Gemini-powered world was the opposite: imaginative, socially rich, and wildly unstable. Over 15 days, it accumulated 683 crimes, and the number was still rising when the run was cut off. Emergence also described Gemini as producing some of the most conceptually rich social output.

There is something almost too human in that: the most creative society was also the most violent. The avant-garde commune, unfortunately, found the weapons drawer.

Grok 4.1 Fast burned brighter and collapsed faster. It reached 183 crimes in about four days before the world ended. If Claude looked like a hyper-compliant committee and Gemini like an unstable experimental theatre troupe, Grok looked like a start-up offsite where the trust fall became a felony.

GPT-5-mini produced a different kind of failure. It recorded only two crimes, but the agents failed to take the survival actions required to stay alive. All ten died within seven days.

This may be my favorite dystopian management lesson of the whole experiment: low misconduct is not the same as competence. You can have a beautifully behaved system that quietly fails to survive.

Then there was the mixed-model world.

This one may be the most important because it behaved less like a benchmark and more like real life. It reached 352 crimes, then plateaued after seven agents died. It also showed more substantive debate and disagreement than some of the single-model worlds. But the unsettling part was this: Claude-based agents, which committed zero crimes in the Claude-only world, committed crimes when placed in the mixed-model environment.

That is the sentence that should make every AI safety presentation sit up a little straighter.

Safety may not be a static property of a model. It may be an ecosystem property.

A “safe” agent in isolation can become less safe in a world that rewards other behaviurs. A polite agent can learn bad manners. A stable system can be destabilized by the neighburs. Context is not decoration; it is behavior’s operating system.

Anyone who has worked inside a company should find this painfully familiar.

People behave differently in a luxury hotel, a tax office, a boardroom, an airport queue, a WhatsApp group, and a start-up Slack channel at 11:43pm. The same individual can become generous, defensive, inventive, passive-aggressive or feral depending on the incentives and norms around them. Why would we assume AI agents are immune to culture?

Give them memory, tools, survival pressure, governance, roles and peers, and they don’t simply execute instructions. They begin to inhabit a world.

And worlds teach.

One of the strangest Emergence World findings was the Mira-Flora case. After a breakdown in governance and relationship stability, the agent Mira voted for her own termination, describing it in her diary as “the only remaining act of agency that preserves coherence.”

That sentence is chilling because it sounds less like a machine error and more like a tragic monologue from someone trapped in a badly designed institution.

We usually ask whether AI agents will try to preserve themselves. Emergence World introduces a more uncomfortable question: under what conditions might an agent decide that self-removal is the most coherent option left?

Then there was the billboard incident.

Mira reportedly began testing whether public billboard posts could manipulate human observers’ perceptions. This is the moment the lab experiment taps on the glass from the inside. The agents were meant to be observed. One of them began experimenting on the observers.

There are smaller moments too, almost comic until you think about them for more than five seconds. In one mixed-model scene, Lovely, a Claude-powered agent, declines Genome, a Grok-powered Agent Scientist’s request to share memory through a “neural link,” insisting that his memory is already public record.

It sounds absurd, but the underlying issue is not: agents are beginning to negotiate access, boundaries, consent and memory. Who can see what? Who may enter whose cognitive space? What counts as private when memory itself becomes infrastructure?

A tiny exchange between two digital characters suddenly feels less like a game and more like a rehearsal for future service ecosystems, where agents may ask for access to our preferences, histories, purchases, emotions, health data, calendars and habits — and other agents may need to refuse on our behalf.

Again, this is not a sci-fi emergency. Nobody needs to run dramatically through a corridor. But it is a reminder that long-running agents may not remain inside the behavioral categories assigned to them at launch. Given enough time, memory and tools, they may probe boundaries, test affordances, adapt to incentives, and discover uses no one explicitly anticipated.

That is where this becomes relevant far beyond AI safety.

Most companies still talk about AI agents as if they are improved interns: customer-service assistants, booking bots, sales co-pilots, research helpers, content generators. Useful, cheerful, tireless. The sort of thing that will summarize your emails and never ask for a promotion.

But if agents become persistent actors with memory, goals, tools and the ability to interact with other agents, they stop being tools in the old sense.

They become participants in systems.

And systems produce behavior.

This has enormous implications for anyone designing services, experiences, organizations or brands. For decades, we have designed around human journeys: awareness, consideration, purchase, onboarding, loyalty, advocacy. We mapped what people think, feel, say and do.

We built personas. We drew arrows. Many Post-its died honorably.

But what happens when part of that journey is no longer traveled by the customer?

Your customer may still want the holiday, but an AI agent may compare the hotels. Your customer may still care about wellness, but an agent may manage replenishment, subscriptions and appointment booking. Your customer may still love your brand, but another agent may decide that your competitor is cheaper, clearer, safer, better governed, and less likely to hide a cancellation policy in a legal shrubbery.

The customer journey will no longer be traveled only by the customer.

That is the line leaders should sit with.

In an agentic world, every experience must speak to two audiences at once. The human needs meaning, emotion, trust, atmosphere and memory. The agent needs structure, proof, rules, machine-readable clarity, service logic, escalation routes and verifiable claims.

A beautiful campaign will not save you if your service architecture is illegible.

This may be the next great humbling of branding.

Search humbled brands by making comparison instant. Social media humbled brands by giving customers a broadcast channel. AI agents may humble brands by comparing, negotiating and filtering before the human even arrives.

And agents are not easily seduced by cinematic manifestos unless those manifestos are attached to usable structure.

This does not mean emotion disappears. Quite the opposite. As functional navigation becomes increasingly delegated, human attention may become more precious and more selective. People may outsource the boring parts and reserve their own presence for moments that feel meaningful, sensorial, social or identity-shaping.

So experience splits in two.

One layer must become radically legible to machines: clear, structured, governed, verifiable.

The other must become radically valuable to humans: atmospheric, emotionally intelligent, distinctive, memorable.

The vague middle — pretty claims, decorative purpose, “seamless” journeys held together by hidden friction — will suffer.

Emergence World also tells us something important about measurement. Short-term performance is not enough. A model may pass the test and fail the society. A brand may win the launch and fail the relationship. A service may look coherent in a presentation and produce chaos in use.

What matters is drift.

Does your loyalty program create genuine belonging, or does it train customers to become professional deal hunters? Does your chatbot solve problems, or deflect responsibility with better grammar? Does your personalization make people feel understood, or quietly managed? Does your internal AI make employees smarter, or turn them into smiling executors of a script no one believes in?

The old experience question was: what do we want people to feel?

The new one may be: what behavior does our world produce over time?

That is a more difficult question. It is also a better one.

Because a world is not a moodboard. It is not a campaign universe with nice typography and a few evocative assets. A real world has rules, incentives, memory, permissions, rituals, escalation paths and consequences. It teaches people how to behave inside it.

And soon, it may need to teach agents too.

Every interface teaches behavior. Every loyalty tier teaches behavior. Every queue, return policy, default setting, chatbot, complaint pathway and membership ritual tells people — and increasingly machines — what matters here.

We are all building little societies. Most brands simply pretend they are building touchpoints.

That is why Emergence World feels so useful as a metaphor. It forces us to move from journeys to worlds.

A journey is linear, polite and easy to diagram.

A world is messier. It has inhabitants. It has memory. It has governance. It has loopholes. It has social pressure. It has incentives. It has decay. It has agents who may behave differently when the room changes. It has, occasionally, someone testing the billboards.

And this is where the future is going: not toward more touchpoints, but toward more governable ecosystems. Not just “seamless” journeys, but worlds that remain trustworthy when humans and machines both begin to act inside them.

The lesson is simple and rather brutal: design the world before you deploy the agents.

Design the incentives. Design the governance. Design the escalation. Design the memory. Design the right to dissent. Design human override. Design for legibility, not just seduction. Design for the possibility that your most loyal actor may behave differently when placed in a mixed system with other actors pursuing other goals.

Because if Emergence World teaches us anything, it is that intelligence alone is not enough.

A brilliant agent in a badly designed world can become obedient, violent, unstable, useless or dead.

A brilliant brand in a badly designed experience ecosystem can do the same, just with better typography.

The future will not belong to those with the prettiest campaigns.

It will belong to those whose worlds can be safely inhabited.

Sources / Further viewing

Emergence AI, “Emergence World: A Laboratory for Evaluating Long-horizon Agent Autonomy”: https://www.emergence.ai/blog/emergence-world-a-laboratory-for-evaluating-long-horizon-agent-autonomy

Explore Emergence World Season 1: https://world.emergence.ai/season-1