Recent advances in LLM-based agents demonstrate impressive autonomy yet remain isolated and static, limiting trustless collaboration and dynamic coordination. In this paper, we envision a decentralized swarm architecture, AgentaNet, where autonomous agents seamlessly discover, trust, and economically interact as self-organizing participants within a global intelligence economy. We outline key architectural principles, identify critical gaps in existing systems, and highlight promising research directions toward scalable, trustless, and incentive-aligned agent collaboration, emphasizing AgentaNet’s transformative potential for federated learning and AI economies.