SwarmWorld: Stigmergic technological evolution in societies of language-model agents
SwarmWorld is a multi-agent system where initially homogeneous LLM agents self-organize without assigned roles or recipes into evolving technological societies. Agents explore a spatial environment, process resources, test materials, construct persistent artifacts, and write executable controllers evaluated by a deterministic simulator under unseen disturbances after the agents are removed. Shared societies develop broader, more resilient technological portfolios than a strong best-of-N isolated-search baseline, although isolated search remains competitive for the strongest artifact.
SwarmWorld demonstrates that decentralized language-model agents can coordinate through a shared environment to build functional technologies and outperform independent search in portfolio breadth and resilience. The system splits cognition from consequence: agents propose architectures and controllers within fixed action and material schemas, while the simulated world determines function. This approach enables collective intelligence to emerge from local actions accumulating into durable social organization.
The key technical contribution is the separation of agent cognition from environmental consequence, allowing stigmergic coordination through persistent artifacts. Agents write executable controllers evaluated by a deterministic simulator under unseen disturbances, ensuring robustness. The finding that shared societies develop broader, more resilient technological portfolios than best-of-N isolated search suggests that collective exploration can mitigate individual search limitations, though isolated search remains competitive for the single strongest artifact.
This research indicates a shift from direct conversation or centralized workflows in multi-agent systems toward decentralized, environment-mediated coordination. Such approaches could enable more scalable and robust AI systems for complex problem-solving, where agents contribute to shared artifacts without explicit role assignment. The competitive performance of isolated search for peak performance suggests hybrid strategies may be optimal in practice.
The ability to generate diverse and resilient technological portfolios through decentralized agent societies could reduce reliance on expensive centralized search or human-designed workflows. This may lower costs for complex engineering and optimization tasks, though the current research is foundational and not yet productized.
Observable next signals include follow-up work on scaling SwarmWorld to larger agent populations, integration with real-world robotic or simulation environments, and comparisons with other multi-agent coordination paradigms. Potential applications in automated design, scientific discovery, and collective robotics may emerge if the approach proves generalizable beyond the current simulation.