Agentic Variation Operators

Agentic Variation Operators (AVO) is a software scaffolding built by NVIDIA that wraps around an AI model. If the AI model itself is the 'brain,' this scaffolding is the skeleton that lets that brain actually move its hands and feet to get work done. What tools the agent can use, how it remembers what it has done before, and how it decides its next move when it's stuck — all of this is determined by the scaffolding.

Just as the same actor's performance can look completely different depending on the director and script they're given, the same AI model can produce very different results depending on what scaffolding it's wrapped in. AVO specifically includes a piece that acts like a 'director' — when the agent starts circling the same spot or repeating a path it has already taken, this component nudges it back on course.

AVO isn't a new commercial product from NVIDIA but an experimental structure released for research purposes. When this scaffolding was applied to Anthropic's Claude Opus 5, its score on a game-based benchmark that requires figuring out the rules without instructions jumped from 30% to 100%.

Articles use it like: "When AVO was applied to Claude Opus 5, its ARC-AGI-3 score jumped from 30% to 100%." A common misunderstanding is thinking AVO is a new AI model — in reality it's just a software structure that wraps around an existing model, and it isn't a commercial product NVIDIA sells.