Nvidia research shifts the AI focus from models to scaffolding

The findings suggest that the industry has spent too much time treating AI models as self-contained brains. According to Adel El Hallack, Nvidia’s vice president of product for its AI unit, an agent is far more than a simple API call. It is a complex ecosystem consisting of the model, the runtime environment, and a set of tools known as the harness. This scaffolding manages memory, context, and feedback—the essential elements for long-horizon tasks that require decision-making over extended periods.

Nvidia’s approach relied on a custom framework called Agentic Variation Operators (AVO). The key innovation involved a 'supervisor' agent—a secondary layer that acts like a CEO, nudging the primary model away from dead ends or repetitive loops. While competitors like OpenAI have struggled with the ARC-AGI-3 benchmark, achieving only marginal gains through minor tweaks, Nvidia’s structural approach suggests that the path to reliable AI lies in better management of the model rather than just scaling parameters.

This shift in perspective mirrors broader industry trends, including recent research from Databricks, which found that the choice of harness can double the operational costs of a project regardless of the model used. By advocating for an open agent stack, Nvidia is signaling that the future of reliable, secure AI depends on giving developers more control over the infrastructure, runtime, and tools that govern how these models interact with the world.