Agent Lightning from Microsoft Research Asia separates agent execution from model training, enabling developers to add reinforcement learning to any existing agent with virtually zero code changes. Works with LangChain, OpenAI Agents SDK, AutoGen, CrewAI, or standalone agents.
Its LightningRL algorithm uses hierarchical credit assignment: after task completion, it scores each LLM request's contribution and feeds step-level rewards into RL algorithms like PPO or GRPO. This lets agents learn from experience rather than relying solely on prompt engineering.
Beyond RL, it supports Automatic Prompt Optimization and Supervised Fine-tuning. With 14K+ GitHub stars and top-50 Microsoft project status, it serves teams needing agents that improve task completion rates over deployment cycles.