2026-10-07 Changes confirmed medium confidence

Microsoft Open-Sources Agent Lightning 1.0 for Real-Harness Agent Training

The 3,500-line framework connects deployed agent harnesses to reinforcement learning through an API proxy and Kubernetes control plane, with benchmark gains still based on Microsoft's own experiment.

Edited by Tyronne Panaino

Microsoft Research Asia open-sourced Agent Lightning 1.0 on October 7, introducing a reinforcement-learning control plane designed to train an agent through the same harness used in deployment. The Microsoft Research release describes a roughly 3,500-line system built around an API gateway, a rollout controller and a customized trainer.

The practical change is architectural. Instead of rebuilding an agent's tool loop inside a training framework, developers can point the harness's model endpoint at an OpenAI-compatible proxy. Agent Lightning then records the model calls and associates them with rollouts while the original harness continues to manage context, tools and execution. That matters to teams whose deployed agents already depend on substantial orchestration outside the model.

Training the harness that will actually run

Microsoft calls the approach Harnessed Agentic RL. Traditional agent reinforcement learning often assumes the training framework owns the environment loop: it receives an action, returns an observation and turns the full exchange into a token trajectory. Modern coding agents complicate that assumption because their harnesses may summarize context, launch subagents, call tools and maintain their own dependencies.

Agent Lightning leaves that surrounding system in place. Its gateway observes requests and responses through the proxy. The rollout controller starts and manages agent executions as local processes or standard Kubernetes jobs. The trainer, built on the verl framework, waits for rollouts, collects samples and uses an adapter to assemble training data. Microsoft says this can run on self-managed clusters, cloud Kubernetes or local infrastructure without requiring a paid commercial sandbox service.

Four problems the framework has to manage

Using a real harness creates a less tidy training record than a single model-environment loop. Microsoft identifies four specific problems: token boundaries can change when text is tokenized again; context summarization or subagents can split one rollout into several samples; sample-level advantage calculations and loss normalization can overweight rollouts that produce more samples; and the training backend must schedule variable workloads before their final sizes are known.

The release addresses those issues with rollout-level accounting and a control plane that separates agent execution from training. It also introduces what Microsoft calls Collocated Async RL, where rollouts and model updates share the same GPUs. In the reported experiment, this scheduling method produced about twice the end-to-end speed of synchronous reinforcement learning while using fewer GPUs than a conventional asynchronous arrangement. That is a first-party result, not a general performance guarantee.

The coding result needs independent testing

Microsoft tested an end-to-end pipeline using SWE-smith, mini-SWE-agent and Qwen3.5-9B. The company reports that reinforcement learning on about 6,000 training samples raised Pass@1 on SWE-bench Verified from 41.8% to 56.4%, an increase of 14.6 percentage points. It also says rollout-level advantage calculation combined with rollout-level normalization delivered better validation reward and steadier policy entropy than sample-level handling in its experiment.

Those measurements show the intended design working in one documented coding-agent setup. They do not establish how Agent Lightning performs across other harnesses, tasks, models or infrastructure, and the fetched release does not provide an independent reproduction, a total-cost comparison or a broad production reliability record. The small codebase may make the system easier to inspect, but line count alone is not evidence of operational maturity.

What to watch next

The most useful next evidence would be third-party reproductions that keep the deployed harness unchanged, publish their workload and compute configuration, and compare training stability, cost and task quality against other agentic RL systems. Reports from teams using different harnesses would also test whether the proxy boundary remains practical when agents have long contexts, nested subagents or heterogeneous tools.

Status

Confirmed. Microsoft has published and open-sourced Agent Lightning 1.0 and documented its architecture and one coding-agent experiment. Internal confidence is medium because the performance and efficiency claims come from Microsoft Research and were not independently reproduced in the fetched evidence.