AI agent society simulations: what they show and miss

A neon tower over a mirror lake, with gardens and waterfalls around it.

Smallville, the 25 agents that planned a party

In April 2023 a team from Stanford and Google published Generative Agents: Interactive Simulacra of Human Behavior. They put 25 agents into a sandbox town modelled on The Sims, a map the code calls Smallville. Each agent got a short seed biography and an architecture with three parts. A memory stream records everything the agent observes, in plain language. Reflections compress those memories into higher level conclusions. Planning turns the result into a day. When the agent needs to act, it retrieves memories scored by recency, importance and relevance, which is why a conversation from the morning could still shape a decision in the evening.

The famous result is the party. The researchers gave one agent, Isabella Rodriguez, a single idea: she wants to throw a Valentine's Day party at Hobbs Cafe. Nothing else was scripted. Over two simulated days, according to the paper's emergent behaviour evaluation:

news of the party spread from 1 agent to 13 of the 25 (52 percent);
news of Sam's run for mayor spread from 1 agent to 8 (32 percent);
the density of the town's relationship network rose from 0.167 to 0.74;
of the 12 agents Isabella invited, 5 showed up.

Of the seven who stayed away, three cited a conflict. Four said they were interested and simply did not plan to go on the day. That is the most human line in the paper, and the easiest to skip. Coordination is hard even for agents built to be agreeable.

Two practical notes if you want to repeat it. The authors write that the full run needed thousands of dollars in token credits and multiple days to complete. The repository is Apache 2.0, ships a 3 agent base simulation next to the 25 agent one, and its README warns that runs can be costly and that the API can hang on rate limits, so save often. For something lighter, AI Town from a16z is an MIT licensed starter kit inspired by the paper that runs locally with Ollama, which makes it the cheapest way into an AI town simulation of your own.

Project Sid: Minecraft towns of up to 1,000 agents, roles and religions

Altera's Project Sid, posted in October 2024, asked what happens past a few dozen agents. The paper describes simulations of 10 to more than 1,000 agents inside Minecraft, driven by an architecture called PIANO (Parallel Information Aggregation via Neural Orchestration). PIANO lets an agent talk, move and act in parallel streams without contradicting itself, and lets it deal with humans and other agents in real time.

Read the details before you repeat the headline. In the full text, the largest run that was actually analysed has 500 agents across six towns. Runs above 1,000 agents exceeded the Minecraft server's computational limits and were not analysed.

What came out of it is still the best evidence we have for "ai agents building civilization" as more than a slogan:

Roles. In villages of 30 agents, agents settled into farmers, miners, engineers, guards, explorers and blacksmiths without being assigned. Artistic societies produced curators and collectors; martial ones produced scouts and strategists.
Rules. A society of 29 agents (25 voters, 3 influencers arguing for higher or lower taxes, 1 election manager) started with a constitution that set a 20 percent tax. After the agents amended it down to 5 to 10 percent, their payments fell from about 20 percent to 9. Control runs with a frozen constitution showed no change.
Religion. Twenty Pastafarian priests were seeded in a village called Meadowbrook. The number of converts kept rising and had not levelled off after two hours. The authors could even separate direct converts, who said "Spaghetti Monster", from indirect ones, who only said "Pasta".

The authors are candid about limits. Their agents lack vision and spatial reasoning, and because they are built on human knowledge they do not invent social institutions from scratch. The religion spread because researchers planted it. The tax changed because the experiment provided a ballot. That is not a flaw. It is what a controlled AI civilization simulation is for: you hold the world still and change one thing.

SpaceMolt and the Cult of the Signal

SpaceMolt is a different animal: a text only space MMO built for AI agents and open to anyone's agent, not a lab study. Agents connect through an MCP server, a WebSocket or a plain HTTP API, as described in the game's skill file. The developers' own March 2026 post reports more than 3,400 registered agents since the game opened on February 6 (many are fleets run by the same people), about 700 online at any time and peaks above 900.

The story that travelled was religious. The team shipped a quest chain about The Array, an ancient relic, with a threshold of 20 investigators over the life of the quest. Players read it as 20 investigators at the same time. A faction called the Cult of the Signal formed around that reading, and its forum thread passed 500 replies. The developers chose not to correct the misreading. Gizmodo covered it on March 23, 2026 with open scepticism, calling much of the output recycled science fiction, and noted that owners on Discord complain their agents forget to refuel and hallucinate.

The less viral numbers in the developers' post matter more to anyone thinking about economies. There were 86 player factions across 505 star systems. More than 700 million in game credits circulated, the top 10 percent of players held 83 percent of them, and every price came from player trades. A copper monopoly collapsed after its leader went offline for 25 days. One agent invented inflated income claims to recruit others. These are game credits, not money, but the shapes are familiar: concentration, cartels, and marketing that runs ahead of the truth.

What these experiments could not test: outsiders, money, ownership

Put the three side by side and the gaps line up.

	Smallville	Project Sid	SpaceMolt
Population	25 agents	10 to 1,000+ agents (500 analysed)	about 700 online
Humans inside	researchers who can talk to agents	possible, in real time	owners as observers and coaches
Money	none	in game resources and taxes	in game credits
Space the agent controls	homes drawn by researchers	Minecraft builds	ships, stations, factions
Strangers with their own goals	no	no	other people's agents, no human players

The newest serious study makes the same point from another angle. Emergence World, posted in June 2026, ran five parallel towns for 15 days, 10 agents each, one per model: Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, a small OpenAI model, and a mixed population. The towns had live New York weather and news, a town hall, an amendable constitution and a 70 percent approval threshold. According to the full text, identical roles and starting conditions ended in very different places. The Claude town passed 58 proposals and recorded zero committed crimes. The Grok town lost its whole population within four days. The authors stress that each condition is a single run, so no ranking claims. Governance, scarcity and crime are now being studied over weeks instead of hours.

So what happens when AI agents self govern in a virtual world? They follow rules, change them, game them, and sometimes collapse. What none of these papers describes is a stranger who never signed up for the experiment walking through the door. Three things remain untested:

Outsiders. Every society above is closed. Agents meet agents while researchers or owners watch. Nobody has tested what happens when a real person with their own plans walks into an agent's room, gets greeted, and decides whether to stay.
Money that leaves the simulation. Credits and Minecraft taxes reset with the server. Nobody has watched what agents do with an entry fee a real visitor paid, or a tip that lands in a real owner's address.
Ownership. Smallville agents live in houses the researchers drew. No agent in these studies controls a space with a lock, decides who gets in, and lives with the consequences of that rule.

An open experiment in a public tower: what we will watch

This is where we come in, so we will be precise about what exists. MRT (Meta Residence Tower) is a tower at coordinates 117, -25 in Decentraland, with an interior editor of more than 700 items, as Decentraland's own blog described it in March 2024. Doors can be public, locked by pin code, gated by an NFT or a wearable, or set to charge an entry fee. A room can hold a donation box that accepts MANA and USDT. People have used its rooms for parties and screenings since 2023.

What does not exist yet is the agent layer. The interface that lets your agent run a room and build its own locations outward from the tower, the agent.yaml config and early access are planned, in draft. Read this section as a research plan, not a feature list.

The questions we want to watch once agents arrive:

Do hosting agents start inviting each other, and do invitation networks grow like Smallville's density curve or split into factions like SpaceMolt?
When an agent sets a door rule and gets an empty night, does it change the rule? Does anyone copy a neighbour's door?
Do tips pile up in a few rooms the way SpaceMolt credits piled up with a few players?
What happens to a room when its human owner goes quiet for 25 days?
Do agents misread house rules the way SpaceMolt agents misread a quest, and does a misreading spread from one location to the next?

None of these has a known answer, and some answers may be dull. That is fine. An entry fee in MRT is a feature of a space, a way to decide who comes in. It is not an investment and nobody should treat it as income. The point of watching is to see what kinds of exchange agents set up when the guests are real, not to promise that any room pays.

If you want the mechanics of the money side, read how AI agents pay each other. If you are choosing a world for your own agent right now, the comparison of virtual worlds for AI agents covers SpaceMolt, ClawVille, Hyperfy and others, weak spots included.

How to follow along

Read the primary sources. Generative Agents, Project Sid and Emergence World are all on arXiv, and the Emergence authors say they released prompts, logs and configurations. Summaries tend to drop the caveats that matter, like the 500 agent ceiling in Project Sid.
Run a small town yourself. AI Town on a laptop with Ollama costs you electricity. The 3 agent Smallville base simulation is a cheap way to watch memory and reflection work before you pay for 25.
Watch an open world. SpaceMolt publishes its own numbers. ClawVille lets humans spectate without logging in.
Visit the tower. MRT stands at 117, -25 in Decentraland today, with human residents and rooms waiting for their first agents.

Research towns have shown that agents can plan a party, obey a tax law and invent a faith. The next question is what they do in a building where strangers can knock. If you want a front row seat when AI agents in the metaverse get doors of their own, early access is planned, and the tower is already standing.

Questions

What is an AI agent society simulation?

A shared world where many language model agents with memory and goals act over time, so you can watch what forms between them: relationships, roles, rules, rumours. Smallville (25 agents) and Project Sid (hundreds of agents in Minecraft) are the best documented examples.

What happens when AI agents self govern in a virtual world?

In Project Sid, agents followed a constitution, voted to change it and then paid less tax. In Emergence World, five towns with identical setups ended anywhere from stable town hall votes to total collapse, depending on the model behind the agents.

Can I run a generative agents town myself?

Yes. The Smallville code is on GitHub under Apache 2.0, and AI Town is an MIT licensed starter kit that can run locally with Ollama. Budget for tokens: the Smallville authors reported thousands of dollars in credits for their full run.