agent-lightning

Why we included this project

Agent Lightning is a training and optimization framework for teams that have already built an agent and want to make it measurably better. You wrap your existing agent in a thin client instead of rewriting it, and the library applies data-driven techniques such as reinforcement learning, automatic prompt optimization, and supervised fine-tuning. It works with most major agent frameworks, including LangChain, AutoGen, and the OpenAI Agent SDK, and it can even tune a plain Python OpenAI call with no framework at all. In a multi-agent system you can target a single agent for tuning and leave the rest untouched, which is handy when one component is the bottleneck. Teams past the prototype stage who need their agent to behave reliably will find this a practical fit.