verl is an open-source reinforcement learning (RL) training framework designed specifically for post-training large language models (LLMs). It supports agentic RL training with features such as server-based asynchronous rollout, multi-turn conversations, and tool calls within an agent framework.