Microsoft's Agent Lightning trains AI agents inside the harness they ship with

The coding result is a before-and-after on one model and one benchmark, and the figures are Microsoft's own. The effect is sizeable. At 41.8% Pass@1, Qwen3.5-9B left 58.2% of SWE-bench Verified tasks unsolved, and after training it left 43.6% ^[8][12]. Training removed about a quarter of the model's failures ^[13].

The design is what makes the result interesting. RL systems such as verl, AReaL and slime were built on the assumption that the trainer owns the agent's loop: the model acts, the environment returns an observation, the observation joins the context, and the rollout becomes one continuous token trajectory ^[1]. Production agents carry their own machinery. OpenHands, mini-SWE-agent and OpenCode each bring their own context management, tool protocols and execution logic ^[2]. Rebuilding one inside a trainer is expensive, Microsoft argues, and the copy may not behave like the original ^[3].

Agent Lightning leaves the loop where it is. It sits between agent and model as a proxy. The team points the agent's model endpoint at it, and it records each call while the harness code stays unchanged ^[10].

Not owning the loop moves the difficulty to the trainer. It sees only a series of request and response pairs, so a single rollout may arrive split into a variable number of parts, one of four challenges the post sets out for training on real harnesses ^[11]. I think that is the right place for the difficulty. It gets solved once, in a codebase of about 3,500 lines that a team can read and change ^[6], instead of in every team's reimplementation of its own agent.

For an in-house fine-tune, the post takes two items off the bill: the rebuild and the commercial sandbox ^[7][3]. The data requirement is small too, about 6,000 samples from an open-sourced dataset ^[9]. The post does not report the GPU hours or hardware behind the run, or how the same model would score if trained on a rebuilt loop with the same data. In my view, compute is the line that decides whether a small task-specific model pays for itself, and it is the one still unpriced. The missing comparison matters for a different reason. The experiment shows this pipeline improves the model; the claim that training on the real harness is the cause rests on the design argument.