Microsoft Research has introduced Agent Lightning v1.0, a lightweight, open-source framework designed to improve AI agents through reinforcement learning. Built with approximately 3,500 lines of code, it trains agents using their existing deployment harnesses without rebuilding execution workflows.
The framework supports Kubernetes-based training, an OpenAI-compatible model proxy and efficient GPU utilization. In testing, Agent Lightning improved Qwen3.5-9B performance on SWE-bench Verified from 41.8% to 56.4%, using approximately 6,000 training samples.
Its approach helps developers reduce training complexity, reuse existing infrastructure and improve coding agents while maintaining consistency between training and production environments across different AI applications.
