Recursive Agent Optimization

RAO-trained recursive agents outperformed flat single-agent baselines across the three main benchmarks. RAO treats recursive execution as a trainable capability rather than just an inference-time wrapper. RAO trains one shared policy at every node of a recursively generated execution tree. Recursive Agent Optimization trains language-model agents that can recursively spawn copies of themselves for delegated subtasks. RAO appears most helpful for tasks that are difficult, long-horizon, context-constrained, decomposable, or parallelizable.