The rapid progress of large language models (LLMs) has gained considerable attention for their universal capability of conducting reasoning, decision-making across diverse domains. These advances have revolutionized natural language understanding tasks, particularly in code generation, contributing to the prosperity of developing autonomous agents for end-to-end programming in software engineering applications. Despite these recent successes, their application to the design of federated learning systems (FL) remains nascent. In this position paper, we advocate for an agentic FL paradigm that harnesses cooperating task-specialized LLM agents to automate the entire FL lifecycle. We outline a four‑stage workflow in which planning, coding, and optimizing agents iteratively generate, refine, and validate FL strategies under a human‑inspired development process. We emphasize open research directions to advance multi-agent FL systems that adaptively configure and manage real‑world FL deployments.