Agent Lightning is a Microsoft open-source Python library for training AI agents via reinforcement learning, automatic prompt optimization, and supervised fine-tuning.

It integrates with any agent framework (LangChain, AutoGen, CrewAI, OpenAI Agent SDK, or raw OpenAI) by inserting a tracer that collects prompts, tool calls, and rewards as structured spans without modifying existing agent code. A LightningStore routes those spans to a selected algorithm that outputs refined prompts or policy weights, while a Trainer orchestrates dataset streaming, resource updates, and inference engine refresh.