Microsoft Research outlined Agent Lightning v1.0 on October 7, describing a way to train AI agents while retaining the software that manages their tools, context and execution.[1]
The framework uses a model proxy to connect an existing agent harness to reinforcement learning. Its design aims to reduce the need to recreate agent behavior inside a separate training system. Microsoft describes a compact implementation and support for running agents as Kubernetes jobs.[1]
In the accompanying research, the authors report that training Qwen3.5-9B with approximately 6,000 examples increased SWE-bench Verified performance from 41.8% to 56.4%, a gain of 14.6 percentage points. Those figures describe the reported experimental setup, not a guaranteed improvement for other models or agents.[2]
The paper was first submitted to arXiv on August 18. The October article is a subsequent official overview, rather than the framework’s first public disclosure.[2]
For developers, its relevance is the possibility of evaluating learning in the same orchestration system used for deployment. Results in other environments will require separate testing.
An agent harness is the surrounding software that decides which tools to call, how to manage context and when to continue a task. The paper argues that this surrounding system should participate directly in training, because rebuilding it elsewhere can introduce differences between training and deployment.[2]
That choice creates additional research challenges. A task can generate several model calls, and the authors examine how those calls become training samples and how rewards and losses are handled without giving undue weight to tasks that produce more samples.[2]
Microsoft's overview also describes three main components: an API gateway, a controller that manages agent execution, and a trainer. This separation allows execution to run independently of the training process.[1]
For teams considering adoption, the published example provides a starting point for reproduction. Its benchmark gain needs to be assessed alongside the model, task environment and training configuration used in the experiment.
Sources:
[1] Microsoft Research, October 7, 2026: https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/
[2] Research paper, first submitted August 18, 2026: https://arxiv.org/abs/2608.17528