Microsoft Agents Learn Without Disrupting Production Setup

Microsoft has released Agent Lightning v1.0, an open-source reinforcement learning framework that allows AI agents to learn within their existing production setup without disrupting it.

This framework is designed to tackle one of the more challenging problems in AI agent development: the gap between training and deployment.

The deploy-time harness, which manages how the agent interacts with the world, remains intact during training, while a proxy captures everything the trainer needs.

Microsoft achieved impressive results using Agent Lightning's RL training pipeline on the Qwen3.5-9B model, achieving a 14.6-point absolute improvement on SWE-bench Verified, pushing the model's score from 41.8% to 56.4%, with only 6,000 training examples and modest computational resources.