Researchers argue that decentralised AI societies face a coordination problem fundamentally different from ordinary message delivery.
Autonomous agents have limited capacity to read and retain incoming information. Instead of pushing every message directly into an agent’s context, the proposed approach gives each agent an inbox governed by its own selection and retention rules. Simulated societies containing up to 16,000 agents demonstrate how local attention policies influence distributed search, negotiation and information propagation. The researchers also find that partial adoption can sharply reduce network-wide search effectiveness, revealing dependencies between individually autonomous participants.
