Microsoft Research ships Agent Lightning in about 3,500 lines

Watch

Lab claim

On October 7, 2026, Microsoft Research released Agent Lightning v1.0, a training framework of about 3,500 lines that keeps the agent harness used in deployment.

How to read this claim

Measurement — The post does not print a repository address on the page we read. The SWE-bench gain is one example.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — Status on this card: Shipped.

Keep the habit · field card with prices and worked checks

What changed

The week's agent news had been new weights. Agent Lightning is a training framework that leaves the existing harness in place.

What it introduces

A small open training framework whose point is to train against the harness you deploy.

Who benefits Anyone training a coding agent and trying to keep the deployed harness instead of rewriting it for the trainer.

Try this

If you already have a harness, read the proxy step before you rebuild the loop. Treat 41.8 percent to 56.4 percent as Microsoft's result on its example, not as your score.

What we do not know

The post does not print a repository address on the page we read. The SWE-bench gain is one example.

Evidence

Primary Microsoft Research. Reported by Microsoft Research. The SWE-bench gain is Microsoft's, on a pipeline it built.

Independent The SWE-bench Verified scores are Microsoft's. No independent rerun is cited.

Sources

Microsoft Research Primary · October 7, 2026

Still unclear? · Start here · ranks, prices, unknown The post does not print a repository address on the page we read. The SWE-bench gain is one example.

The record

What happened

On October 7, 2026, Microsoft Research published Agent Lightning v1.0. The post says the framework is about 3,500 lines and that the harness used in deployment is the harness that takes part in reinforcement learning, through a proxy in front of the model. In the coding example, Microsoft says about 6,000 training samples raised Qwen3.5-9B from 41.8 percent to 56.4 percent Pass@1 on SWE-bench Verified, a gain of 14.6 points. Agents can run as Kubernetes jobs.

Why it matters

A new model is not the only way an agent changes. This post is about training the model inside the harness that will actually run. The 14.6 point gain is one pipeline, Microsoft's, on SWE-bench Verified. It is not a promise that every harness will move by that amount.

Microsoft says the example used SWE-smith, mini-SWE-agent, and Qwen3.5-9B. Collocated Async RL is described as about twice the end-to-end speed of synchronous RL in their experiments.