In Project Sid, 30 AI agents divided up work in Minecraft in 20 minutes

You launch several AI assistants: one is supposed to study the task, another to check the solution, and a third to carry out the actions. But two agents take on the same work, while nobody handles a necessary step. In Project Sid, the Altera team investigated a similar problem in Minecraft. In an experiment with 30 characters, stable roles emerged within roughly 20 minutes, even though participants had not been assigned professions in advance.

This offers useful insights for anyone building teams of agents. It shows why an assistant needs to know what the others are doing, and why a shared instruction alone may not be enough. Whether the findings transfer from the game to real workflows still needs to be tested.

How agents found work that needed doing

Participants received a shared task: survive together and build an efficient settlement. They started with identical personality traits and information about important locations: a forest, a resource-rich cave, farmland, and a central square. From there, each chose its actions independently.

Imagine two residents near the forest. One is already chopping down trees. The second notices this and can switch to gathering food or mining ore. This kind of observation of neighbors helped distribute the work. New social goals emerged every 5–10 seconds, allowing participants to adjust their plans as they worked.

The researchers identified six professions: farmers, miners, engineers, guards, explorers, and blacksmiths. There is an important caveat: these labels were assigned after the experiment. GPT-4o was used to classify the characters’ goals. The agents themselves did not go through a job assignment process, although their actions matched the identified roles.

The most revealing test involved disabling the social perception modules. Without them, stable specializations largely stopped forming: characters switched roles more often and performed similar work. This supports the conclusion that information about other participants’ actions helped coordination in this particular environment.

What enabled collaboration

The characters’ behavior was driven by the PIANO architecture, which had ten modules. These handled memory, goals, social signals, conversations, and in-game actions. Several processes could run in parallel: a character did not have to stop observing the world entirely while thinking through its next step.

A cognitive controller gathered information from the modules and coordinated key decisions. This addresses a fairly everyday situation: an agent promises to bring a neighbor some iron, but another process is already sending it off to work on the farm. The architecture is intended to help align speech, plans, and actions; the design alone does not guarantee absolute reliability.

In separate runs, the team studied changes to rules. They prepared a constitution, a tax system, and a voting procedure in advance for 29 agents. Following a campaign to lower taxes in one run, the rate fell from 20% to 5–10%, while actual contributions amounted to around 9% of inventory. The agents were able to change their behavior in response to a collective decision within the predefined mechanism.

What you can apply to your own product

Large-scale experiments reached more than a thousand participants, but the server struggled and characters got stuck. Cultural processes were studied in detail with 500 agents. However, the spread of Pastafarianism was assessed through words used in conversations: mentioning a religion does not, on its own, demonstrate belief.

Navigation and collaborative construction remained challenging. The models also relied on human training data and background information prepared by the researchers. The observed behavior cannot automatically be treated as the independent invention of social institutions or an accurate model of humans.

For developers, this suggests a working hypothesis: teams of assistants should have access to up-to-date information about other agents’ actions, shared memory, and agreed decisions. You can test it on a familiar task by measuring how much work agents duplicate with and without this information exchange. Project Sid provides a reason to run that experiment, but it does not yet offer a ready-made recipe for a reliable AI team.