Harness Scaling for Generalizable Agentic Reinforcement Learning

The paper proposes Harness Scaling, extending agentic reinforcement learning from a single fixed execution harness to diverse execution configurations. Harnesses are organized along five components (system prompt, tool interface, observation, context management, workflow), with a compact training subset selected for balanced factorial coverage. Across multiple benchmarks, model scales, and RL algorithms, the authors report improved generalization to unseen white-box configurations, and transfer of models trained this way to unseen external black-box harnesses such as Claude Code.

Key takeaways

• Harness Scaling consistently improves generalization to unseen white-box harness configurations across models, benchmarks, and RL algorithms.
• Structured harness selection and locality-aware training are reported as more effective than simply increasing the number of training configurations.
• Models trained across diverse white-box harnesses transfer to unseen black-box harnesses such as Claude Code, reportedly outperforming direct black-box RL training.

Why it matters

Agent harnesses shape how models use tools and manage context, so RL trained through only one harness may not generalize. Organizing harness diversity could make agentic RL more robust across deployment setups.

Read with care

Source claims not independently checked; the supplied text is the authors' promotional post and abstract, not the full paper results.