Microsoft Bridges Training-Harness Gap with Agent Lightning v1.0

Microsoft has released Agent Lightning v1.0, an infrastructure concept for agent optimization that addresses structural challenges in post-training LLM-based agents during reinforcement learning processes.

The framework aims to bridge the disconnect between training engines and production harnesses by ensuring the production harness is engaged from the start, overseeing infrastructure services and agent interactions throughout both initial training and subsequent reinforcement learning actions.

Using Agent Lightning v1.0 on modest compute with 6K training examples, reinforcement learning improved Qwen3.5-9B on OpenAI's SWE-bench Verified benchmark by an absolute 14.6-point gain, from 41.8% to 56.4%

Software engineer Md Rashedul Hasan notes that training through the real harness keeps semantics intact and reduces train-serve mismatch.

Hasan believes Microsoft has successfully named the paradigm, kept the core framework small, and shipped a concrete coding-agent pipeline with open data and scripts.