AVO
Nvidia's research agent system, built for tasks that run for days rather than minutes.
AVO stands for Agentic Variation Operators and it is Nvidia’s answer to a specific problem: language models are good at single answers and bad at week-long projects. AVO wraps a model in a system that keeps notes between steps, calls tools, runs its own tests, and has a second component watching the overall direction that can step in when the agent starts going in circles.
It was built to optimise GPU kernels, the small pieces of code that decide how fast a graphics card runs, where it worked unattended for seven days. Nvidia later pointed the same system at the ARC-AGI-3 reasoning benchmark without changing the agent itself, which is the interesting part: what transferred was not knowledge about the task but the machinery for sticking with one.