Large language model (LLM) agents are increasingly deployed in private environments, where they gather valuable experience from task execution, tool usage, feedback, and local knowledge. This experience is scattered across organizations and cannot be shared directly due to privacy and proprietary constraints. Conventional federated learning only synchronizes model parameters and therefore cannot accommodate the broader spectrum of agent capabilities such as memory, tools, rewards, skills, and structured knowledge.
The paper introduces Federated Agent Optimization (FAO), which studies how distributed agents can collaboratively improve through controlled information exchange while keeping raw data, full trajectories, and private knowledge local. FAO is formulated as a multi‑objective optimization problem that balances agent utility, privacy leakage, and communication cost. To address this, the optimization space is organized into five dimensions: policy, memory, tool use, reward, and structured knowledge/skills.
The authors further describe how private experience can be abstracted, protected, aggregated, and adapted into transferable capabilities, providing a unified view that enables agents to benefit from each other without sharing raw experiences. Finally, key challenges are identified—including privacy metrics, cross‑domain knowledge alignment, communication efficiency, and trustworthiness assessment—and several promising research directions are outlined, such as differential privacy mechanisms, federated meta‑learning, and interpretable capability mapping, to advance trustworthy federated agent systems.
Review