Microsoft introduces Agent Lightning v1.0, a framework for training AI agents without breaking their production setup

Microsoft just shipped a framework that tackles one of the more annoying problems in AI agent development: the gap between how you train an agent and how you actually deploy it. Agent Lightning v1.0, released on August 17, is an open-source reinforcement learning framework that lets AI agents learn within the same harness they already use in production.

What harnessed agentic RL actually means

The framework introduces a paradigm Microsoft calls “harnessed agentic RL.” In traditional reinforcement learning setups for AI agents, you typically need to reconstruct the agent’s entire interaction loop inside a training environment. Agent Lightning flips the script. The deploy-time harness, meaning the existing infrastructure that manages how an agent interacts with the world, continues to govern the entire interaction loop during training. The trainer component only observes sequences of requests and responses from the underlying large language model. It never touches the harness itself.

In practical terms, this means developers can apply reinforcement learning to their agents with zero changes to pre-existing code, tools, context, control flow, or environments. The agent talks to what it thinks is a normal LLM endpoint, but behind the scenes, Agent Lightning has slotted in a proxy that captures everything the trainer needs.

The entire core framework is roughly 3,500 lines of Python.

The benchmark results are hard to ignore

Microsoft backed up the release with evaluation results on SWE-bench Verified, a benchmark that measures an AI agent’s ability to resolve real-world GitHub issues from popular open-source Python repositories.

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, pushing the model’s score from 41.8% to 56.4%. That jump came from only 6,000 training examples and what Microsoft described as modest computational resources.

The release includes complete workflows, training scripts, and a reproducible pipeline specifically designed for coding agents.

Technical architecture and what’s under the hood

Agent Lightning v1.0 integrates two key pieces of infrastructure: verl, a framework for RL training of LLMs, and vLLM, a high-throughput inference engine. The Trainer component sits at the center, managing the learning process while proxy mechanisms handle the communication between the agent’s existing harness and the training infrastructure.

The framework also includes advanced support for rollouts using Kubernetes, which matters for teams running agent training at scale.

This version builds on foundational work from the earlier iteration of Agent Lightning, which came out of Microsoft Research Asia’s Shanghai lab. The v1.0 release represents a significant rewrite of that original codebase.

The code is available on GitHub under the MIT license.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.