FL-MAESTRO Optimizes Federated Learning with LLM Agents

This research introduces FL-MAESTRO, a novel multi-agent orchestrator designed to enhance the efficiency of Federated Learning (FL) in environments with resource-constrained and volatile edge devices. In FL, the communication network and client availability are dynamic, requiring real-time decisions on which clients participate, how resources are allocated, and how their updates are aggregated. FL-MAESTRO employs three specialized Large Language Model (LLM) agents, each responsible for one of these decision dimensions: communication topology, per-client resource allocation, and the aggregation rule. A central coordinator synthesizes their analyses into a single decision, which is then validated by a non-LLM feasibility check before execution. A key innovation is the orchestrator's ability to consume a predicted-failure list from the server, allowing it to proactively exclude clients whose updates are unlikely to be aggregated. This mechanism drastically reduces wasted energy, a major issue in classical FL on unstable edge networks. The system demonstrated matching accuracy with leading energy-aware baselines on a CIFAR-10 benchmark while virtually eliminating wasted round energy.