By Jon Markman
Publication Date: 2026-08-24 15:36:00
NVIDIA components are displayed at the GTC Paris NVIDIA at the VivaTech technology startups and innovation fair at the Paris Expo Porte de Versailles, in Paris on June 12, 2025. The VivaTech fair opened in Paris on June 11, 2025 in the presence of the French Minister for Digital Technologies, before welcoming a number of tech stars and the French President against a backdrop of trade tensions between Europe and the United States. (Photo by Thomas SAMSON / AFP via Getty Images)
AFP via Getty Images
Nvidia took a frontier AI model that completes about 30 percent of one of the field’s hardest benchmarks and got a perfect score out of it.
The model was Anthropic’s Claude Opus 5, and nothing inside it changed: no retraining, no fine-tuning, not one adjusted weight. Everything that improved sat outside the model, in a software system Nvidia calls AVO that manages the model’s memory, plans its next moves, and watches for mistakes. Better software pulled three times more performance out of a model that already exists.
What AVO Actually Is
The software layer between a model and a task is called a harness, and Nvidia’s researchers describe its job in one clean passage:
A frontier language model is only one component of an AI agent. The surrounding agent system—often called a harness—determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and sustains progress over long-running tasks.
AVO, short for Agentic Variation…
