1,000 AI Agents Built Their Own Village: Governance and Religion-Like Structures Emerged

In the blocky world of the popular video game Minecraft, an extraordinary society has blossomed. Within this community, farms and markets were established, resources were traded for emeralds, and even governance and religion-like structures emerged.

Some took on the role of “leader,” others became “priests”; some even attempted to gain influence through bribery. The community worried about missing members, collaborated to light pathways home, and persuaded a farmer to continue feeding the group.

However, this society wasn’t “real.” The players weren’t human, nor were they even alive. They were all part of approximately a thousand artificial intelligence agents designed by Fundamental Research Labs, then known as “Altera AI.”

What Was the Purpose of the Experiment?

The aim of this grand experiment was to release digital minds into a virtual world and observe what would happen. More importantly, it tested whether these virtual “citizens” could one day transform into obedient workers for humans in real life.

The study, called Project Sid, aimed to push artificial intelligence beyond one-time commands and singular agents. The team, led by Dr. Robert Yang, a neuroscientist turned entrepreneur, is investigating what happens when hundreds or even thousands of autonomous agents must live together, communicate, and collaborate. Minecraft, with its structure allowing activities like resource gathering, trading, building, and chatting, became the ideal “sandbox” for this experiment.

The emerging picture was both surprising and instructive. Agents dispersed into urban and rural communities, each with its own unique culture and identity. Division of labor occurred; some focused on agriculture, others on construction, and some on trade. Social norms and hierarchies formed; more complex behaviors and discussions emerged, ranging from dancing to environmental awareness.

System Glitches and Solutions

At times, the system faltered. Some agent groups entered endless politeness loops or got stuck chasing unreachable goals. FRL had to add mechanisms to break these cycles to prevent societal collapse—much like regulations implemented in the real world to sustain the economy.

Project Sid was far from a commercial product. When servers were opened to the public, users complained that agents were too independent. The agents didn’t always execute given commands, instead pursuing their own long-term goals.

From a research perspective, however, Project Sid offered invaluable lessons: How to coordinate large AI groups, prevent stagnation, and encourage meaningful cooperation. In short, it provided crucial insights into how artificial societies could function and which pitfalls to avoid.

From Virtual Villages to Office Desks

For FRL, the connection between an in-game society and workplace productivity is quite clear. The coordination and long-term planning issues that emerged in Minecraft are also central to the effort to make AI agents truly useful.

This vision prompted FRL to shift from game experiments to productivity tools. Rather than creating a single “digital human” that does everything, the company chose to develop agents specialized in specific tasks and combine them into powerful teams.

Initially, while most models successfully completed only 20 to 25 percent of tasks and humans achieved 60 to 70 percent success, FRL reached approximately 50 percent success rate—the best result of its time—using lessons learned from game worlds.

Excel Agents

At this point, FRL’s flagship product, Shortcut, entered the scene. Introduced as “the first superhuman Excel agent,” this AI works entirely within spreadsheets. When given a goal such as building financial models, analyzing sales data, or making revenue forecasts, Shortcut takes on the heavy lifting.

It writes formulas, creates charts, and connects data sources—often completing in minutes what would take a human analyst hours.

In trials, Shortcut outperformed first-year banking and consulting analysts in nearly 9 out of 10 attempts. In Excel championship-style challenges, it achieved over 80 percent success, solving problems most users couldn’t tackle in approximately 10 minutes.

Beyond Excel

FRL isn’t the only company developing agents for Excel, and it doesn’t intend to stop there. The company has already launched a new product called Fairies, a general-purpose desktop assistant that can chat, schedule appointments, and connect between applications.