Microsoft Research details Agent Lightning v1.0, a small framework for training agents in their real harnesses

Microsoft Research Asia has published a blog post describing Agent Lightning v1.0, an open-source reinforcement-learning framework rebuilt around what the researchers call Harnessed Agentic RL. In this approach, the same harness an agent uses in deployment takes part directly in training, rather than being reimplemented inside the training framework.

The researchers say the reimplementation step is costly and means the agent being trained is "not quite the agent that gets deployed". Earlier systems such as verl, AReaL and slime were built so the training framework owned the interaction loop. Agent Lightning instead places an LLM proxy between the agent and the model. According to Microsoft, pointing an agent's existing model endpoint at the proxy is usually enough to connect it, leaving harness code unchanged.

Microsoft says the framework is about 3,500 lines of code, split across an API gateway, a rollout controller and a trainer built on verl. It says agents can run as standard Kubernetes jobs on self-managed clusters, cloud Kubernetes or local infrastructure, without paid commercial sandbox services. The post also describes a Collocated Async RL mode in which rollouts and model updates share the same GPUs, and reports about a 2x end-to-end speedup over synchronous RL.

The headline result is from a coding-agent pipeline built on SWE-smith, mini-SWE-agent and Qwen3.5-9B. The researchers report that Pass@1 on SWE-bench Verified rose from 41.8% to 56.4%, a gain of 14.6 percentage points, using about 6,000 training samples.

The post is open about difficulties. Because the trainer sees only request and response pairs, one rollout can become a variable number of training samples. Retokenizing text can shift token boundaries so adjacent calls cannot always be merged, and counting samples equally can give extra weight to rollouts that happen to produce more of them. The researchers say rollout-level advantage and normalization gave higher validation reward and steadier policy entropy than sample-level handling in their coding experiments.

What is confirmed and what is not: everything above comes from Microsoft Research's own post. The numbers are the researchers' reports, and we have not seen independent replication. A technical report and GitHub project are linked from the post but were not reviewed here, and the timing of the first code release relative to this post is unverified.

For builders, the practical point is the claim that an existing deployment harness can be used for training without being rebuilt. Whether that closes the gap between trained and deployed agents beyond this one benchmark remains to be shown. This is a single-source item.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.