Microsoft Releases Agent Lightning: A New AI Framework that Enables Reinforcement Learning (RL)-based Training of LLMs for Any AI Agent

Microsoft's open-source framework, Agent Lightning, enables reinforcement learning (RL) training for large language models (LLMs) and multi-agent systems without requiring modifications to existing agent architectures. By formalizing agents as partially observable Markov decision processes (POMDPs), the framework extracts clean RL transitions from real agent tracesfocusing solely on policy calls, inputs, outputs, and rewardsthereby simplifying the conversion of complex multi-step interactions into standard RL training data. Agent Lightning introduces LightningRL, a hierarchical method that converts multi-step agent runs into single-turn RL transitions compatible with common trainers like PPO