Agent Lightning
Reinforcement learning framework for training LLM agents in real harnesses
What it is
Agent Lightning trains the language model behind an AI agent with reinforcement learning while the agent keeps running its own tools, context, and control flow. A gateway sits between the agent and the model to capture each call as training data, a controller starts agents on a local machine or as Kubernetes Jobs, and a trainer built on verl and vLLM updates the policy. ML researchers and engineers use it to improve agents such as coding, search, and math agents against benchmarks, with worked examples covering SWE-bench, Search-R1, and GSM8K.