Microsoft Research Asia Open-Sources Agent Lightning v1.0: A Lightweight Agentic RL Framework

Summary:Microsoft Research Asia has released Agent Lightning v1.0, a lightweight agentic reinforcement learning (RL) framework designed to streamline AI agent training. By integrating seamlessly with real deployment harnesses, it eliminates the need to rebuild agents within the training framework. The framework, weighing in at around 3,500 lines of code, features native Kubernetes support and an efficient data training recipe, boosting Qwen3.5-9B's performance by 14.6 percentage points on the SWE-bench

---

Microsoft Releases Agent Lightning v1.0: A Lightweight Agentic RL Framework

Microsoft Research Asia has open-sourced Agent Lightning v1.0, a lightweight agentic reinforcement learning (RL) framework designed to address key challenges in traditional RL training. Here are the main features and technical highlights of the framework:

Key Features

1. Lightweight Design: The entire framework consists of approximately 3,500 lines of code, making it easy to understand, modify, and extend.
2. Seamless Integration with Real Harnesses: Through an LLM proxy, Agent Lightning v1.0 can directly interact with existing agent harnesses, eliminating the need to rebuild agents for training.
3. Native Kubernetes Support: Agents run as standard Kubernetes jobs, supporting self-managed clusters, cloud Kubernetes, and local infrastructure.
4. Efficient Data Training Recipe: The end-to-end coding agent training pipeline, based on Qwen3.5-9B, boosted the model's performance from 41.8% to 56.4% on the SWE-bench Verified benchmark, using only about 6,000 training samples.

Technical Highlights

• Harnessed Agentic RL Paradigm: Agent Lightning v1.0 introduces the Harnessed Agentic RL paradigm, allowing agents to participate directly in reinforcement learning with the same harness used in deployment, avoiding the complexity of rebuilding agents within the training framework.
• Collocated Async RL: The framework features Collocated Async RL, enabling rollout and model updates to share the same set of GPUs, thereby improving resource utilization and reducing GPU requirements.
• Kubernetes Native Support: By leveraging Kubernetes jobs, Agent Lightning v1.0 can efficiently manage large-scale agent training tasks, lowering costs and enhancing scalability.

Industry Impact and Developer Recommendations

Agent Lightning v1.0 opens new possibilities for AI agent training, particularly in scenarios requiring integration with complex toolchains. Developers can leverage this framework to quickly build and optimize agents without worrying about inconsistencies in agent behavior during training. Additionally, native Kubernetes support makes large-scale training tasks more efficient and cost-effective.

For developers interested in trying out the framework, here are some recommendations:

• Familiarize Yourself with Kubernetes: Since Agent Lightning v1.0 has native Kubernetes support, understanding the basics of Kubernetes will help you make the most of the framework.
• Optimize Training Data: While the framework supports efficient training methods, high-quality training data remains crucial for improving model performance.
• Explore Collocated Async RL: Experiment with Collocated Async RL to further enhance training efficiency and resource utilization.

Conclusion

The release of Agent Lightning v1.0 marks a significant milestone in the field of agentic reinforcement learning. Its lightweight design and efficient training methods provide a powerful tool for the development and optimization of AI agents.

---

Source: Microsoft Research AI (2026-10-07)

— END —

Tags: #Microsoft Research Follow #Agent Lightning Follow #Reinforcement Learning Follow #Kubernetes Follow #Open Source Framework Follow