Researchers from Microsoft Research Asia release Agent Lightning v1.0 to train agents using their own harnesses

Researchers from Microsoft Research Asia have released the completely redesigned Agent Lightning v1.0 framework as open-source. Its Harnessed Agentic RL approach uses the same agent harness for reinforcement learning as for deployment. The agent retains its own context management, tool use and interaction control; the training system records model requests and responses through an LLM proxy. According to the authors, redirecting the model API endpoint to this proxy is usually sufficient.

According to the authors, the new version contains approximately 3 500 lines of code. API Gateway records data for training, Rollout Controller runs agents as local processes or Kubernetes jobs, and Customized Trainer builds on verl. The Collocated Async RL approach allows agent execution and model updates to share the same GPUs. In the authors' experiments, it roughly doubled end-to-end training speed compared with synchronous RL and required fewer GPUs than conventional asynchronous RL.

The authors validated the entire process using SWE-smith data, mini-SWE-agent and Qwen3.5-9B. The training set contained approximately 6 000 samples; the process included preparing the data and environment, as well as safeguards against reward hacking. According to their measurements, RL alone increased the score achieved by Qwen3.5-9B on SWE-bench Verified from 41.8 % to 56.4 %, a gain of 14.6 percentage points.