Microsoft Research releases Agent Lightning v1.0, a lightweight RL framework for real agent harnesses
Agent Lightning v1.0 is a 3,500‑line open‑source framework that trains agents using their deployment harnesses, with native Kubernetes support and a demonstrated coding‑agent performance boost.
Original source published: October 7, 2026
Microsoft Research Asia has open‑sourced Agent Lightning v1.0, a reinforcement‑learning framework built around the Harnessed Agentic RL paradigm. The entire system is about 3,500 lines of code and connects directly to an agent's existing harness via an LLM proxy, eliminating the need to reimplement the agent inside the training loop.
The framework runs agents as standard Kubernetes jobs, avoiding commercial sandbox services, and introduces Collocated Async RL, which the authors report yields roughly a 2× end‑to‑end speedup over synchronous RL while using fewer GPUs. In an end‑to‑end coding‑agent experiment, training Qwen3.5‑9B on about 6,000 samples raised Pass@1 on SWE‑bench Verified from 41.8% to 56.4%, a 14.6‑point gain.