Microsoft Research introduces Agent Lightning v1.0 for RL agent training

Microsoft Research has introduced Agent Lightning v1.0, a lightweight agentic reinforcement learning framework aimed at training AI agents with real harnesses. The framework addresses a common challenge in agent development: tools, context, and decision-making are often handled inside complex agent frameworks, making reinforcement learning difficult to apply directly.

Agent Lightning connects existing agents to RL training so teams can improve their behavior without rebuilding the agents from scratch. The approach is positioned as a way to make reinforcement learning more practical for agent systems that already rely on established tooling and runtime environments.