SwarmWorld: When AI Agents Build Societies Through Shared Tools
This presentation explores SwarmWorld, a system where societies of language-model agents develop persistent technological ecologies through decentralized coordination and physical stigmergy. Unlike typical multi-agent systems that prescribe roles or workflows, SwarmWorld allows identical agents to evolve behavioral differentiation, construct material artifacts with executable controllers, and transmit culture through both communication and physical object modification. The research reveals that shared worlds produce broader, more resilient technological portfolios than isolated agents, though not necessarily the single best artifact, and demonstrates that coordination emerges primarily through encountering and inheriting physical technologies rather than direct communication alone.Script
Most AI agent systems assign roles and prescribe workflows. SwarmWorld flips that premise: it starts with identical agents in one shared world and asks what technological societies emerge when they must explore, fabricate tools, and respond to disturbances entirely on their own.
The simulator enforces a hard separation between what agents claim and what actually works. Agents propose technologies and write controllers, but function is determined by physics, material constraints, and deterministic program execution. An artifact continues operating even after every agent is removed.
Technologies evolve through executable inheritance. Agents fork controller programs, modify instructions, and install them in new artifacts. The deepest lineage spans 12 generations, and under full cultural conditions, about half of all program forks cross author boundaries. This is not metaphorical evolution but content-addressed provenance of working code.
Here is the central finding: shared societies produce broader and more resilient technological portfolios than isolated search, but isolated agents still build the single best-performing artifact. Over 3,200 ticks, independent search achieves an artifact performance of 0.35, while the full-culture society reaches only 0.24. Yet the society generates 7 validated inventions and maintains portfolio resilience of 0.25, nearly double the isolated envelope.
Agents begin identical with no assigned roles, yet behavioral differentiation emerges anyway. By tick 3,200 under full culture, over half the population adopts artifact-centered work, clustering around shared infrastructure. This is not a fixed profession but a dynamic state allocation responding to the evolving technological landscape.
The deepest surprise is that most technology transfer happens not through messages but through physical encounter. Ninety-five percent of first artifact reuse begins when agents simply discover installed objects in the world. The environment itself becomes collective memory, where construction by one agent reshapes the opportunity landscape for all others. Visit EmergentMind.com to explore this research further and create your own videos.
