A unified framework for RL training of language agents, supporting any harness, any sandbox, and one-flag backend switching.