MSR Asia Releases Agent Lightning: A 3,500-Line Agentic RL Framework for Real Harnesses
Original title:Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses
Training agents with reinforcement learning often forces developers to reimplement interaction loops separate from production harnesses. Microsoft Research Asia's Agent Lightning compresses the entire RL control plane into roughly 3,500 lines of code, running real deployment harnesses directly atop Kubernetes without proprietary sandbox dependencies. In an end-to-end coding benchmark using approximately 6,000 samples, the setup lifted Qwen3.5-9B's Pass@1 score on SWE-bench Verified from 41.8% to 56.4%.
Why it's worth reading
It bridges the architectural gap between agent deployment harnesses and RL training setups, demonstrating double-digit benchmark gains using modest code and sample budgets.