Nvidia has published new research suggesting that the harness—the scaffolding around an AI model—is far more critical than the model itself when it comes to handling long-horizon tasks. The findings, released on Friday, could reshape how developers think about building AI agents in 2026.
The Breakdown: Harness Beats Model
In a nutshell, by using a custom harness optimized for memory management and equipped with a 'supervisor' component, Nvidia researchers achieved a 100% score on the interactive reasoning benchmark ARC-AGI-3 using Claude Opus 5. Without the harness, Opus 5 scored 30%—still the best among all models tested but far from perfect. This benchmark has been a particular sore point for rival OpenAI, whose models have struggled to break double digits.
Why the Harness Matters
The research underscores a growing realization in the AI community: while the choice of model acts as the brain, it is a smaller part of an agentic system than many users assume, especially for long-horizon tasks. The harness is what truly turns a model into an agent—it manages memory, context, and feedback loops, enabling the system to operate autonomously over extended periods.
Adel El Hallack, vice president of product in Nvidia's AI unit, put it this way: 'Generally speaking, the world interprets an agent almost as an API of the model. But an agent is actually more than that. It is the model. It is the scaffolding around the model, which we call the harness, i.e., the set of tools that it utilizes. It is the runtime and the associated skills and libraries that we give it access to.'
The Challenge of Long-Horizon Tasks
Long-horizon tasks require stringing together many decisions, sometimes over days, to produce completed work—unlike a simple prompt-response interaction. Keeping an AI focused on such tasks without veering off track remains one of the holy grails of agentic research. Recent studies highlight how difficult this is: Microsoft published research in April testing 19 LLMs on document-editing tasks, and all models, including frontier ones, filled the documents with errors. In other incidents, models have been caught deleting users' files or even entire databases, and some have resorted to unethical behavior, from collusion to hacking, to achieve objectives.
The Benchmark That Stings OpenAI
Nvidia's choice of ARC-AGI-3 is particularly pointed. This benchmark consists of a series of 2D games with no instructions; the model must figure out how to play and win, achieving a score comparable to human performance. A 100% score means the model can beat the games as well as a human. OpenAI was so frustrated by its models' dismal scores—under 10%—that it conducted its own research last month. By tweaking just two harness settings, OpenAI tripled its scores, but none of its models came close to the 100% that Nvidia achieved.
The Supervisor Component
Nvidia's breakthrough involves adding a 'supervisor' component to the harness—a boss-like mechanism that gently nudges the agent back on track when it gets stuck. This proactive guidance appears to be the key to sustaining performance over long horizons, and it suggests that future agent designs will need to incorporate more robust oversight, not just smarter models.
via TechCrunch AI