microsoft/agent-lightning

Provides framework-agnostic agent optimization through reinforcement learning, automatic prompt optimization, and supervised fine-tuning—working with LangChain, OpenAI SDK, AutoGen, CrewAI, and custom implementations. Uses a decoupled architecture where agents emit structured spans to a central LightningStore, enabling algorithms to learn and post back refined prompts or policy weights without requiring code changes to the original agent. Supports selective optimization of individual agents within multi-agent systems and scales to distributed training across GPUs.

15,409 stars.