NVIDIA Skill2Env: RL training from 3.4k agent skills, 4.7% terminal bench gain

NVIDIA researchers compiled 3.4k public agent skills into 8k terminal tasks with automated tests and rubrics. Using just 300 steps of GRPO RL, Qwen3.8-27B improved 4.7 points on Terminal-Bench 2.1. The data pipeline includes plan, diversify, create, and verify stages, costing ~$90k API compute.