Beschreibung
Build an autonomous agent-led reinforcement learning research workflow using agent skills that leverage NVIDIA NeMo RL, NeMo Gym and Brev. This tutorial shows how Codex can configure a GPU environment, run experiments and track results using reusable agent skills for Brev etiquette, session memory, and autoresearch.
This workflow covers two examples:
1. Visual counting with RL: Create a visual star-count environment for Qwen3-VL-2B-Instruct, scale training, and improve reported accuracy from 25.0% to 96.875%.
2. Paper-to-code validation: Demonstrate how an agent can help translate a research paper into code and start validation training, while the researcher remains responsible for goals, budgets, and final judgment.
Tech blog:
Brev Launchable: https://brev.nvidia.com/launchable/deploy?launchableID=env-3ECKNnhrpAkQVmcfgA88kFWW7v7
NeMo RL on GitHub: https://github.com/NVIDIA-NeMo/RL
NeMo Gym on GitHub: https://github.com/NVIDIA-NeMo/Gym
0:00 - Agent-Led Coding and Research
0:21 - What Is Autoresearch?
0:47 - Three Capabilities in the Workflow
1:27 - NVIDIA NeMo RL, NeMo Gym and Brev
1:42 - Three Agent Skills for Codex
2:42 - Setting Up the Brev Environment
4:31 - Paper-to-Code Workflow
6:25 - Goal-Driven Autoresearch and Scaling
8:06 - Key Takeaways
8:39 - Learn More and Try It Yourself