NVIDIA shows coding agents autonomously running research

On July 15, NVIDIA's NVIDIAAI and NVIDIA Developer accounts posted a series of demos for an "agent-led autoresearch" workflow: a researcher gives a coding agent a goal and a time budget, and the agent then sets up the training environment, runs experiments, and completes training and evaluation on its own, with the human mainly supervising the process. What makes the set noteworthy is not a single metric but how it chains automated research, reinforcement-learning training, and managed GPU resources into an executable pipeline aimed at open-source models like Nemotron.

Key details

NVIDIAAI's example task has the agent teach a vision model to count colored stars, combining Autoresearch, NeMo RL, NeMo Gym and Agent Skills. NVIDIAAI also shared a Nemotron Labs practice of running RL autoresearch with Agent Skills, turning "auto-exploration" into a runnable agent workflow. NVIDIA Developer added that on GPU instances Codex can use Agent Skills to auto-configure the environment and run experiments for open-source model experimentation and domain customization.

Managed training and infrastructure

Beyond the general autoresearch flow, NVIDIA Developer showed two managed paths: one combining Brev so Codex auto-configures the GPU environment and runs experiments via agent skills; the other using Prime Intellect to do RL customization on open-source models without standing up your own GPU cluster, demonstrated on Nemotron 3 Nano from cold start through exporting a LoRA adapter.

The posts disclosed no experiment metrics, comparisons, or a concrete release timeline, so the emphasis at this stage appears to be on showing how the method and toolchain fit together rather than headline results.

2026-07-15 ~ 2026-07-15 · 6 related posts