article --summary
The summary
This piece explains how Nvidia worked with Factory to capture reproducible failures from the Droid coding agent, covering problems with tool calls, argument formatting, and recovery from errors. Those traces became post-training data for the Nemotron 3 Ultra open model. The reader will also learn how the PivotRL reinforcement learning method, which focuses on critical decisions rather than matching reference examples, fits into the effort.
Nvidia guarded against overfitting to one harness by varying Droid's configuration and mixing data from other agent environments such as OpenCode and Codex. The article also contrasts this with Hugging Face's approach of training separate LoRA adapters for different harnesses, which lifted a Qwen model's Terminal-Bench 2.1 score from 39.5% to 55.8%. It notes that neither company has published direct comparisons of the performance and compute trade-offs between these two methods, and that Factory has not released an ablation study for its share of the gains.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.