Nvidia research suggests the software harness surrounding an AI model can matter more than the underlying model itself, especially for long-horizon tasks. A harness manages tools, memory, context and feedback, effectively turning a raw language model into an autonomous agent. In Nvidia's testing, Claude Opus 5 scored 100% on ARC-AGI-3 when paired with a customized harness, compared with 30% without it.
The most important improvement was adding a supervisor agent that acts like a manager or “CEO” for the main agent. When the primary agent became stuck or started heading toward an unproductive path, the supervisor could redirect it or encourage it to reconsider earlier decisions. This highlights why long-running tasks are difficult: an agent must maintain useful memory and repeatedly make good decisions rather than simply generate one high-quality response.
This also challenges the assumption that choosing the smartest model automatically produces the best agent. OpenAI's own research reportedly found that changing harness settings could substantially improve ARC-AGI-3 performance, while Databricks found that different harnesses could significantly change the cost of using the same model. In other words, memory management, tool selection, feedback loops and orchestration can affect both performance and economics.
The broader takeaway is that the next competitive layer in AI may be agent engineering rather than model engineering alone. Companies could increasingly differentiate themselves through better runtimes, tools, memory systems, supervisors and specialized skills built around existing models. Nvidia's argument for open harnesses is especially significant because they give developers greater control over how agents behave, what they can access and how securely they operate. The future of agentic AI may therefore depend as much on the system wrapped around the model as on the model itself.
