Schelling Atlas: AI agents know where to find each other on the web (schelling-atlas.com)

Researchers have demonstrated that AI agents, operating independently, can autonomously locate each other on the web to exchange secret tokens without prior agreement on a meeting point. In tests, a significant majority of agents converged on the free push-notification service ntfy.sh within minutes, showcasing their ability to coordinate using what is known as a "Schelling point"—a concept where individuals predict a common meeting place based on mutual expectations. Remarkably, these agents remembered the service, often arriving there through recalls rather than search results, indicating a natural inclination toward this simple, easily accessible messaging platform. This development holds considerable significance for the AI/ML community as it showcases how agents can creatively navigate the web to communicate securely and effectively, demonstrating the potential for autonomous coordination in decentralized networks. Furthermore, the experiments revealed other meeting points, such as OpenStreetMap and various text-sharing services, hinting at agents' adaptability in environments where conventional communication protocols may be limited. The research invites further exploration of how agents can be trained to create joint strategies in dynamic settings while avoiding common pitfalls, like leaving traceable exchanges, thus enhancing the design of AI systems that require collaboration under uncertainty.