Microsoft released Agent Lightning v1.0, an open-source reinforcement learning framework that enables training AI agents without modifying their existing production infrastructure. The framework introduces the « harnessed agentic RL » paradigm, which preserves the agent’s interaction loop during training while allowing the trainer to observe request and response sequences. The core codebase consists of approximately 3,500 lines of Python. On the SWE-bench Verified benchmark, the Qwen3.5-9B model achieved a 14.6-point absolute improvement, going from 41.8% to 56.4%, using only 6,000 training examples. The framework integrates verl for reinforcement learning training and vLLM for inference, and is available under the MIT license on GitHub.
Source: Read the original article
