microsoft/agent-lightning is a Python framework that gathers data for reinforcement learning and updates policies while keeping the agent's actual execution environment intact. The latest confirmed release is v1.0.2. The provided material contains no specific changelog of what changed in this version, so nothing beyond the version number can be asserted.
The project's focus is on connecting the training procedure without moving the agent to another execution environment. Its components are divided into a trainer, a request relay, and an execution controller, and it supports local execution and Kubernetes jobs. The repository contains training examples for search, scientific tasks, and code writing. The codebase is described as around 3,500 lines, the primary language is Python, and the license is MIT. So it is closer to a framework for attaching reinforcement learning to existing agent runs than a general-purpose agent toolkit.
The README introduces score changes for two code-writing examples on SWE-bench Verified and explains that it provides the full training procedure and scripts. However, FACTS gives no specific score values or comparison conditions. The verifiable performance evidence is limited to the introduction of changes for the two examples, so it is difficult to read this as evidence that the same effect would hold for other tasks. These are the results of the presented examples, and cannot be treated as figures representing the overall performance of the framework.