Microsoft Launches Agent Lightning v1.0 for AI Agent Training

Microsoft has released Agent Lightning v1.0, an open-source framework designed to enhance the training of AI agents without disrupting their existing production environments. Launched on August 17, this new tool allows developers to implement reinforcement learning while maintaining their current code, tools, and infrastructure intact.

The framework introduces a concept termed 'harnessed agentic RL,' which enables AI agents to learn within the same operational framework used during deployment. Unlike traditional reinforcement learning methods that require a complete overhaul of the interaction loop, Agent Lightning allows the existing infrastructure to manage the interaction process during training, with the trainer component merely observing the exchanges between the agent and the large language model.

Benchmark results from the SWE-bench Verified indicate a significant performance improvement when using Agent Lightning. The Qwen3.5-9B model achieved a score increase from 41.8% to 56.4% after training with just 6,000 examples, demonstrating the framework's effectiveness with modest computational resources. The release includes comprehensive workflows, training scripts, and a reproducible pipeline tailored for coding agents.

Agent Lightning v1.0 integrates two essential infrastructures: verl, a framework for reinforcement learning training of large language models, and vLLM, a high-throughput inference engine. This version builds upon earlier work from Microsoft Research Asia, representing a substantial rewrite of the original codebase. The code is available on GitHub under the MIT license.

© 2026 KLEA News. All Rights Reserved. This article is provided for informational purposes only. It is not offered or intended to be used as legal, tax, investment, financial, or other advice.

Source: KLEA News