Summary
On August 17, 2026, Microsoft released Agent Lightning v1.0, an MIT-licensed, open-source reinforcement-learning framework that trains and optimizes AI agents without changing their existing code, tools, or environments. A full rewrite of roughly 3,500 lines of core Python, it routes agents through a proxy so training decouples from execution, adds native Kubernetes scale-out, and works with agents built on LangChain, the OpenAI Agents SDK, AutoGen, or from scratch.
What changed
Microsoft published Agent Lightning v1.0 as open source under MIT: a reinforcement-learning framework that trains agents while leaving their code, tools, and runtime unchanged by separating agent execution from the training loop via a proxy, with native Kubernetes support (each rollout becomes a Kubernetes Job). Using its RL pipeline on Qwen3.5-9B, the team reports a 14.6-point absolute gain on SWE-bench Verified, from 41.8% to 56.4%.
Why it matters
It lowers the barrier to systematically improving agents with reinforcement learning, with no rewrite and no framework lock-in, making agent optimization accessible to teams that already run working agents. As an MIT-licensed release from Microsoft, it also seeds an ecosystem around training agents, not just running them.
Evidence excerpt
Agents interact with the model through the Agent Lightning proxy with ZERO changes, while keeping tools, context, control flow, and environments in the loop... Using Agent Lightning's RL training pipeline on the Qwen3.5-9B model, the team achieved a 14.6-point absolute improvement on SWE-bench Verified, from 41.8% to 56.4%. Released under the MIT license.