LEMON: Counterfactual RL for Multi-Agent Orchestration

LEMON (Learning Executable Multi-agent Orchestration via Counterfactual Reinforcement Learning) is a new LLM-based orchestrator that generates executable orchestration specifications for multi-agent systems. It integrates task-specific roles, customized duties, capacity levels, and dependency structures into a single deployable system. Unlike existing approaches that optimize these decisions partially or sequentially, LEMON uses counterfactual reinforcement learning to provide better credit assignment for local orchestration decisions. The system addresses the challenge that LLM-based multi-agent systems depend heavily on orchestration design, including role design, capacity assignment, and dependency construction, which jointly affect solution quality and execution efficiency. The approach is detailed in a paper on arXiv (2605.14483).