Microsoft Research Asia has released Agent Lightning v1.0 and formalized Harnessed Agentic RL: the same agent harness used in deployment participates directly in reinforcement learning instead of being reimplemented inside the trainer. An existing agent can point its model endpoint to an OpenAI-compatible proxy, while the framework records calls, manages rollouts and runs agents as local processes or Kubernetes jobs. The complete control plane is about 3,500 lines. In the reported experiment, roughly 6,000 samples raised Qwen3.5-9B from 41.8% to 56.4% Pass@1 on SWE-bench Verified. This is a research and engineering result, not a guarantee for other models, tasks or production agents; training cost, reward design and safety controls still need separate validation.