# NVIDIA's Skill2Env compiles public Agent Skills into 7,971 terminal tasks

A Codex planner and creator turn each skill into tested terminal workflows, using a rubric taken from that skill's own quality criteria. Outcome-only RL moved Qwen3.8-27B from 49.4% to 54.1% on Terminal-Bench 2.1, while adding the rubric reached 50.1% but won more trajectory comparisons.

Language: en
Time zone: America/Los_Angeles
Current time (UTC, cached as of): 2026-11-10T07:31:00.000Z
Query date (America/Los_Angeles): 2026-11-09

HTML: https://didcodexreset.com/news/ad2ab417d6a84d5bad5e4ed6.html

Content language: en
Localization state: sameLanguage

Source: DAIR.AI · Published 9/23/2026, 13:44:25

NVIDIA's Skill2Env compiles each public Agent Skill into executable terminal tasks for reinforcement learning. A Codex planner reads the SKILL.md bundle, researches related public assets, and splits the skill into workflows. A Codex creator then builds each task with programmatic tests and a behavioral rubric taken from the skill's own quality criteria.

From about 3,400 crawled skills, the pipeline produced 7,971 tasks across 13 domains, with software engineering under a quarter of the corpus. Generating them with GPT-5.6 Sol cost more than $90,000 in API usage. The write-up is the [Skill2Env paper](https://github.com/NVlabs/Skill2Env/blob/main/paper/Skill2Env_arXiv.pdf).

After 300 steps of outcome-only RL, Qwen3.8-27B improved from 49.4% to 54.1% on Terminal-Bench 2.1 and from 33.4% to 37.7% pass@1 on S2EBench, its hand-verified held-out benchmark. Adding the rubric to the reward produced a smaller benchmark gain, to 50.1% on Terminal-Bench 2.1. Given the source SKILL.md, a judge preferred the rubric-trained model's trajectories over the base model's on 73.0% of tasks, compared with 54.5% for the outcome-only model.

Tags: Skill2Env, NVIDIA, Agent Skills, Reinforcement learning, Qwen

[View original post](https://x.com/dair_ai/status/2102857541776707799)