Hundreds of AI agents just built an autonomous civilization in Minecraft

What happens when you drop hundreds of LLM-powered bots into Minecraft with no script and no human intervention?

They don't just wander aimlessly. They form villages, invent distinct specialized jobs, debate local tax laws, and even convert each other to the Church of the Flying Spaghetti Monster.

This isn't sci-fi. It’s Project Sid, a wild simulation from AI startup Altera. Founder Robert Yang quit his gig as a computational neuroscience professor at MIT after reading Stanford’s famous 2023 virtual town paper. Backed by $11 million from a16z and Eric Schmidt’s fund, he set out to see what happens when autonomous agents are pushed to their absolute limits.

Emergent jobs, tax debates, and noodle priests

Altera built modular brains for the bots. Specific LLM-driven components handle speech, social reactions, and long-term planning, while game-engine hooks translate those thoughts into picking up tools or trading.

Starting with smaller groups, weirdly human patterns showed up almost immediately. The bots tracked social cues and developed mutual likeability scores. In one run, an AI chef decided to hand out bigger food rations to villagers who made him feel appreciated.

In a 30-agent run where everyone started with the exact same stats and prompts, roles emerged out of nowhere. Without being told to specialize, some agents spent their days gathering seeds as farmers, guards set up defensive perimeters, and an aspiring artist just walked around picking flowers.

When Yang’s team cranked the server up to its limit—between 500 and 1,000 agents—things got even weirder. After seeding a few bots with the parody religion Pastafarianism, the AI priests successfully converted neighboring agents, who then traveled down the road to proselytize in nearby towns. In another test, agents successfully persuaded each other to vote for or against local tax hikes based on casual conversations.

Fake feelings, real utility

Altera cofounder Andrew Ahn is clear on one point: these agents aren’t conscious. The underlying models simply contain an absurdly detailed statistical map of human culture and social dynamics, which the game environment brings to life.

Still, Yang wants to push way past Minecraft. He plans to bring the tech to Roblox next, with the ultimate dream of building companion agents that live alongside humans in the physical world. Yang even argued that AI could one day learn to genuinely love humans, drawing an analogy to how dogs love their owners.

That take ruffled feathers. Julian Togelius, head of game-testing firm Modl.ai, pointed out that a neural net running on a GPU feels nothing at all. But he conceded an important practical point: if an AI can simulate care so convincingly that it brings comfort to someone, users might not care whether the machine actually feels it.

Why it matters

Project Sid is one of the most compelling proofs yet that multi-agent systems can scale without collapsing into gibberish.

Most current agent frameworks struggle with three or four bots coordinating a simple coding task. Showing that a thousand decentralized LLMs can maintain distinct personalities, follow social rules, and divide labor without a centralized orchestrator is a massive technical milestone. Game studios will immediately eye this for living open-world NPCs that actually live their own lives instead of reading canned dialogue trees.

Beyond entertainment, we think the real power lies in macro-simulation. Economists and policy designers could soon run complex tax, housing, or public health interventions inside a digital sandbox of thousands of agents to spot unintended behavioral consequences before rolling them out in real life.

That said, the ease with which Pastafarianism spread through the towns exposes an uncomfortable flip side. If LLM crowds are this susceptible to peer influence, malicious actors could just as easily seed radicalization, scams, or synchronized panic across agent networks. Keeping synthetic societies stable will be a lot harder than building them.