Harness Scaling for Generalizable Agentic Reinforcement Learning

Harness Scaling for Generalizable Agentic Reinforcement Learning. By Fei Bai and seven others, from Renmin University of China, IQuest Research, and one other institution. Published on the twenty-first of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Agent harnesses have become a central part of modern agentic systems. They shape how a model uses tools, how it manages context, and how it coordinates interaction over a long horizon. A harness only pays off if the model underneath can make good use of it, which motivates reinforcement learning carried out directly through the harness. Existing approaches, though, typically train through one fixed harness. That can produce specialisation, and limited transfer when the harness configuration changes.

Harness Scaling proposes a new scaling dimension for agentic reinforcement learning. Instead of training through a single harness, training is extended across diverse execution configurations. The variety of harnesses a model trains through becomes something to scale.

The recipe has three steps. First, a factorized harness space is constructed, so that a harness is described by a set of factors that can vary. Second, a compact training subset is selected from that space with balanced factorial coverage. Third, cross-harness training is organised to combine two things: global coverage across the space, and controlled variation for the same task. The same task seen through different harnesses is where the controlled part comes from.

Each step answers a different need. The factorized space makes the diversity explicit, so it can be enumerated instead of left to chance. Balanced factorial coverage keeps the training subset compact without letting any factor go unrepresented. And pairing global coverage with controlled variation means the model sees the whole space while also seeing one task under several configurations, where only the harness differs.

Across multiple benchmarks, model scales, and RL algorithms, Harness Scaling consistently improves generalisation to white-box configurations that were not seen during training. The improvement holds across all three of those axes of variation, not in one favourable setting. The stronger result concerns harnesses from outside. Models trained on structured white-box variation transfer effectively to unseen external black-box harnesses, and they outperform models trained through a single complex black-box harness.

The authors' conclusion has two halves. Harness diversity is a meaningful scaling dimension for agentic reinforcement learning. And how that diversity is organised matters for generalising across harnesses: the factorial coverage and the controlled per-task variation are part of the result, not just the number of harnesses. Structured variety across white-box configurations transferred better than one complex black-box harness.