Microsoft has released its agent reinforcement learning (RL) framework, Agent Lightning v1.0, on GitHub.

According to a recent report by The New Stack, the core of Agent Lightning v1.0 is that a production harness, rather than a training engine, handles context construction, tool execution and the agent-environment interaction loop.

The training engine observes only the LLM question-and-answer logs exchanged with the harness. As a result, developers do not need to separately build, inside the training system, the logic for an agent to use tools and interact with its environment.

Microsoft reinforced-trained the Qwen3.5-9B model using 6,000 training data items with what it called "moderate computing resources". It said the SWE-Bench Verified benchmark score rose 14.6 points to 56.4 percent from 41.8 percent.

In traditional approaches, the training engine owns the entire interaction loop, creating issues such as re-tokenisation, sample merging, advantage calculation, loss normalisation and training backend scheduling. To address these issues, Microsoft also provided a data-cleaning pipeline for coding agents and reproducible training scripts. The full framework consists of about 3,500 lines of core Python code and was released under an MIT license.

Rashid Hasan (라시드 하산), a software engineering researcher based in Nebraska, said training through an actual production harness reduces mismatch between training and serving. He said tool protocols, context policies and recovery behaviors remain intact, allowing training outcomes to carry over better into real production environments.

Priyank Jain (프리양크 제인), a data scientist based in Colorado, assessed the approach as an attempt to eliminate the "training-serving gap," one of the oldest and most expensive problems in machine learning. "The real achievement is that what you trained becomes what you actually deployed," he said. He added he was concerned more people could attempt reinforcement learning without properly understanding reward design.

Rhea Banerjee (리아 바네르지), founder of TooledX, gave a positive assessment that the codebase is small enough for infrastructure engineers to read directly and trust. She added that in practice it may be usable only by platform teams that have GPU and Kubernetes clusters. She also pointed out that because changes to the harness are reflected in model weights, the harness itself needs version control.