Microsoft Research releases Agent Lightning v1.0 for training agents with deployment harnesses

Microsoft Research Asia has introduced Agent Lightning v1.0, an open-source framework built around a training approach it calls Harnessed Agentic RL. The method brings the agent harness used in deployment into the reinforcement-learning process, rather than rebuilding the agent inside the training system. The framework is about 3,500 lines of code, which the researchers say makes it easier to understand, modify and extend.

Earlier agentic reinforcement-learning systems generally put the training framework in charge of the agent’s interactions with its environment. Developers often had to recreate the agent’s loop in that framework, a costly step that could leave the trained agent behaving differently from the deployed version. Agent Lightning instead uses an LLM proxy to observe model requests and responses while the existing harness manages the interaction loop.

A builder can point an existing model endpoint to Agent Lightning, with the blog saying the harness code can remain unchanged. The system records the model calls, but a single run may produce a variable number of training samples, complicating how those samples are combined and how training advantages and losses are calculated. The researchers also flag a scheduling issue: the training system learns the sample count and length only after a run finishes, while GPU and parallel-processing configurations are typically set in advance.

Agent Lightning supports running agents as Kubernetes jobs on self-managed clusters, cloud Kubernetes or local infrastructure, without relying on paid commercial sandbox services. In its coding-agent pipeline, the researchers report that Qwen3.5-9B improved from 41.8% to 56.4% Pass@1 on SWE-bench Verified. The gain was 14.6 percentage points, using about 6,000 training samples.

For AI builders, the practical distinction is that reinforcement learning can use an existing deployment harness instead of requiring its interaction loop to be recreated inside the training framework. That may help keep the agent being trained closer to the one that will be deployed, which is the mismatch the researchers say their approach addresses. The design still has technical complications: converting text context back into tokens can shift token boundaries and prevent adjacent calls from being merged. The researchers also warn that counting samples equally can give more influence to runs that happen to produce more samples.