Microsoft Releases Agent Lightning to Train AI Agents in Their Existing Setup

Developers can now train AI agents while keeping the software that manages their tools, context, and execution intact. Microsoft Research Asia announced and open-sourced Agent Lightning v1.0 on October 7, 2026. The framework routes model calls through a proxy, allowing the agent’s deployment setup to participate directly in reinforcement learning rather than be rebuilt for training.

An agent harness is the surrounding software that controls tools, context, and the steps an agent takes. Reinforcement learning teaches systems through trial and error, using rewards and penalties. Microsoft’s premise is that rebuilding a harness inside a training framework is expensive—and can leave developers training something that behaves differently from the deployed agent.

Agent Lightning instead inserts an OpenAI-compatible model proxy between the harness and the model. Microsoft says changing the model endpoint is usually enough to connect an existing harness. The framework contains roughly 3,500 lines of code across three components: an API gateway, a rollout controller, and a customized trainer.

The gateway records prompts, responses, and the model probabilities needed for training. The controller launches agent runs, while the trainer collects their resulting samples. Agent execution stays separate from the trainer, instead of handing the training framework control of the agent’s interaction loop.

Simply averaging over those samples can give extra weight to runs that produce more of them, even when that count reflects harness behavior. Other challenges include text-to-token conversion and scheduling variable workloads on fixed GPU resources.

The coding pipeline combines SWE-smith, mini-SWE-agent, and Qwen3.5-9B, with data cleaning, environment construction, and safeguards against gaming the reward. Microsoft reports Qwen3.5-9B improving from 41.8% to 56.4% Pass@1 on SWE-bench Verified—a 14.6-percentage-point gain—using about 6,000 training samples. That result applies to the specified pipeline, not every supported harness.

Agents run as standard Kubernetes jobs on self-managed or cloud clusters, or on local systems, without requiring paid commercial sandbox services. Agent Lightning also lets agent runs and model updates share GPUs. When enough runs finish, the gateway pauses new requests, waits for active requests to complete, and resumes agent work after the update. Microsoft reports about twice the end-to-end speed of synchronous training in its experiments.