Federated Learning (FL) faces significant challenges in enabling seamless knowledge sharing across decentralized services, each with distinct data structures, privacy constraints, and domain specific goals. In this paper, we present AgentFL, a novel agent mesh architecture that leverages agentic AI powered by LLMs to overcome these challenges. AgentFL represents a transformative leap in autonomous FL, enabling diverse AI services to collaboratively share knowledge without sacrificing their distinct objectives or data
sovereignty. At the core of AgentFL is a four-agent architecture, each built on LLMs, designed to answer a key question in modern AI: how can specialized services—with unique data, goals, and privacy
needs—effectively learn from one another? The Orchestration Agent manages and coordinates the system’s operations. The Registration Agent handles client onboarding, access control, and compliance.
The Monitoring Agent continuously tracks system performance and privacy metrics, proactively flagging potential risks. The Strategic Agent develops training plans and knowledge-sharing strategies, relaying decisions via the Orchestration Agent. Through the collaborative interplay of these four agents, AgentFL achieves robust, privacy preserving knowledge transfer across diverse domains, setting the
foundation for scalable, collaborative AI ecosystems.