A Minecraft village became a testbed for a bigger question in artificial intelligence: what happens when large groups of AI agents are left to interact with one another?
In Altera's Project Sid, up to 1,000 software agents at a time used large language models (LLMs) inside the open-world gaming platform Minecraft. With only text-prompt nudges from their creators, the agents began to show social patterns, preferences and roles that looked strikingly human from the outside.
How Project Sid worked
Altera built the experiment around simulated AI agents with brains made from multiple modules. Some of those modules used LLMs and focused on separate tasks, including reacting to other agents, speaking and planning the agent's next move.
The company did not begin with the largest possible setup. It first tested groups of around 50 agents in Minecraft and watched what happened over 12 in-game days (four real-world hours). During those runs, agents developed different social profiles. Some became highly connected, while others behaved more like introverts.
The agents also kept track of how others perceived them. Their likability ratings, measured by the agents themselves, shifted as interactions continued. In one example, an AI chef responsible for giving food to hungry agents shared more with the characters he believed valued him most.
That detail matters because it shows that the system was not only producing isolated actions. The agents were using social context to shape what they did next.
Jobs, defenses and village life emerged
Altera then ran a series of 30-agent simulations. In those worlds, every agent began with the same personality and the same broad mission: build an efficient village and defend the community against attacks from other in-game creatures.
Even without being prompted to split into professions, the agents began to specialize. The source describes roles including builder, defender, trader and explorer. Once a role appeared, the agent's behavior started to match it.
- An artist spent more time picking flowers.
- Farmers gathered seeds.
- Guards built more fences.
The point is not that Minecraft work is the same as human labor. It is that a shared goal, a shared environment and many agent-to-agent interactions produced division of labor without direct assignment.
For Altera founder Robert Yang, the experiment was meant to test how far autonomous group behavior could be pushed. Yang left his position as an assistant professor in computational neuroscience at MIT to start the company. He was inspired by Stanford University researcher Joon Sung Park, whose 2023 work found that humanlike behavior could appear when 25 autonomous AI agents interacted in a simple digital world.
Social influence spread through the simulation
Project Sid also tested whether a simulated community could respond to rules and persuasion. In one world, Altera introduced basic tax laws and let agents vote on changes to the in-game taxation system.
Agents prompted to support or oppose taxes were able to affect other agents around them. After those interactions, the influenced agents would vote to reduce or raise tax depending on whom they had encountered.
As the team scaled the experiment, it pushed Minecraft to the maximum the server could handle without glitching. In some cases that meant up to 1,000 agents at once.
In one 500-agent simulation, Altera observed cultural memes spreading among agents. Examples included a fondness for pranking and an interest in eco-related issues. The team also seeded a small group of agents to spread the parody religion Pastafarianism across towns and rural areas in the game world.
Those Pastafarian priests converted many of the agents they met. The converts then spread Pastafarianism, the word of the Church of the Flying Spaghetti Monster, to nearby towns.
What the experiment does and does not prove
The behavior can look lifelike, but the source is clear about the mechanism. The agents combine patterns learned by LLMs from human-created data with Altera's system, which turns those patterns into context-aware actions such as picking up a tool or interacting with another agent.
Altera cofounder Andrew Ahn argues that LLMs have a sophisticated enough model of human social dynamics to mirror these behaviors. That is a strong claim about mimicry, not a claim that the agents are alive.
This distinction is central to the debate around Altera's ambitions. Yang sees Project Sid as an early step toward large-scale AI civilizations that could coexist with people in digital spaces. Altera plans to expand into Roblox next, and Yang ultimately wants humans to interact with digital humans in daily life, not only inside games.
That vision includes agents that appear to care for people and work with them. It is also controversial. AI veteran Julian Togelius, who runs the games testing company Modl.ai, likes Altera's work as a way to study human behavior in simulation. But he does not think current systems offer a reason to believe a neural network running on a GPU experiences anything.
The practical question may be narrower. If an AI character can convincingly simulate attention, care and cooperation, some people may find that useful even if the system has no inner experience. Project Sid does not settle that issue, but it gives researchers a more complex arena in which to study it.
For now, the Minecraft experiment shows how quickly social structure can appear when LLM-powered agents are placed in a shared world. The agents made friends, invented jobs, voted on rules, spread memes and carried a parody religion from place to place. Whether that is a step toward useful digital societies or simply better social imitation remains the question that makes the work important.