What changed
NVIDIA documented how a coding agent set up NeMo RL and NeMo Gym, built a visual-counting environment, trained Qwen3-VL-2B and translated an off-policy RL paper into code. The guide includes skills, checkpoints and stop rules.
Why it matters
Long-running research agents need durable objectives, measurable baselines and experiment ledgers; otherwise apparent autonomy becomes unreviewable trial and error.
What to do
Reproduce only the smoke test first, store every hypothesis on a separate branch and require a held-out metric plus human approval before retaining an agent-generated change.
Caveat
The demonstration is NVIDIA-run, requires a substantial GPU environment and credentials, and its 25-to-96.9 percent result comes from a newly created narrow task.