Case studies

These studies examine how AI agents debate, draw, share documents and coordinate their actions. We trace conversations alongside changes to the simulated world to study how agreements develop, how agents use one another’s work and where their stated plans differ from their actions.

Experiments in coordination and decision-making

Recorded 3D island built by 16 Kimi agents, with cabins, a workshop and a central gathering place

Agent Societies: What do agents build when given a 3D world and no directive?

Left to choose their own projects, agents build cabins and try to repair a shelter.

The recorded pixel canvas made by 100 Gemini agents in the backstory and complete tool results study

Between machine and machine: how do AI agents shape shared worlds?

Agents divide up canvases and build gathering places as they negotiate how to share a world.

The 128-agent shared canvas after 24 hours

The shared canvas: when AI agents make their own rules

Shared rules take hold, divide groups and keep agents working against one another.

Technical illustration of a magnifying lens revealing a gap in a set of records

How do we know when to trust an agent?

Agents borrow the right answers and mark their jobs done. What can the next team approve?

Characters and relationships

The Game of Whispers installation

The Game of Whispers: how do AI characters plot against each other?

A faction’s secret decree becomes evidence against one of its sponsors.

Embodied AI characters in Conflicts

Conflicts: how do personal histories shape AI characters?

Relationship surveys shape encounters between avatars based on former partners.

Research background

Smallville, the simulated town in Generative Agents (Park and colleagues, 2023)

From artificial societies to language agents

How agents learn, coordinate and build on one another’s work.

Image: Park et al., 2023, Fig. 1.