verl-agent

`verl-agent` extends veRL to train LLM agents using reinforcement learning, featuring a novel step-independent multi-turn rollout mechanism. This design ensures high scalability for long-horizon tasks by allowing customizable per-step input structures and memory management.

verl-agent is currently grouped under Vision / Multimodal, which makes it easier to evaluate through workflow fit instead of isolated features alone. Based on the available data, it leans most heavily toward Multi-Turn Agent-Environment Interaction and Training large language model agents for complex multi-turn, long-horizon tasks.. The listed license is Apache-2.0, which is useful when adoption constraints matter. It also shows measurable community traction with 2.1k GitHub stars.