How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo

Autonomous Agents Tackle End-to-End Reinforcement Learning Research

A technical demonstration has shown that an advanced coding agent, Codex with GPT 5.5, can successfully automate an entire reinforcement learning (RL) research workflow. This process includes setting up the full software stack on NVIDIA hardware, orchestrating experiments, and even implementing novel algorithms directly from research papers. The development is notable for its use of specialized agent 'skills' to manage complex, long-running machine learning campaigns, establishing a framework for reproducible and robust research that moves beyond simple, single-shot coding tasks.

The workflow was executed on an NVIDIA Brev GPU instance, leveraging the NVIDIA NeMo RL and NVIDIA NeMo Gym frameworks. Key to the agent's success was the integration of three complementary skills: 'Brev-etiquette' for system management, 'session-memory' for state persistence, and 'autoresearch' to manage the experimental loop. In a practical test, the agent autonomously created a novel visual counting environment and trained a Qwen3-VL-2B-Instruct model, substantially improving its accuracy from a 25% baseline to 96.9%.

• Agent Platform: Codex with GPT 5.5
• Core Frameworks: NVIDIA NeMo RL, NVIDIA NeMo Gym
• Hardware: NVIDIA L40S 48 GB GPU on a Brev instance
• Agent Skills: Brev-etiquette (system hygiene), Session-memory (state persistence), Autoresearch (experiment management)
• Demonstrated Result: Improved a vision-language model's accuracy from 25.0% to 96.9% on a custom-generated task.

This agent-led approach allows research teams to offload the time-consuming, iterative work of setup, debugging, and experimentation. It frees up human researchers to concentrate on high-level strategic decisions, such as setting research goals, reviewing milestones, and interpreting final results. By containing the workflow within a controlled repository, this methodology provides a practical path for organizations to build specialized domain agents while maintaining direct control over their data, intellectual property, and model training processes.

The true advance demonstrated here is not just the automation of research tasks, but the codification of operational knowledge into reusable 'skills,' which enables AI agents to execute complex, long-running scientific workflows with the reliability and reproducibility required for serious research.