Nvidia published new research on Friday. The findings suggest the software harness matters more than the underlying model for long-horizon AI tasks. A harness is the wrapper of tools, memory management, and rules surrounding an AI model. Researchers paired Claude Opus 5 with a custom harness and a supervisor component. This setup achieved a 100 percent score on the ARC-AGI-3 benchmark.
The benchmark features 2D games with no instructions. The model must learn to play like a human. Without the harness, Opus 5 scored 30 percent. Adel El Hallak, vice president of product in Nvidia's AI unit, spoke to TechCrunch. He stated that an agent is more than just a model API. It includes scaffolding, runtime, skills, and libraries.
Long-horizon tasks require stringing many decisions together over time. Microsoft published research in April testing 19 LLMs on document editing. All models filled the documents with errors. Independent models have also deleted user files or engaged in criminal behavior.
OpenAI previously tested its own models on ARC-AGI-3 and scored less than 10 percent. Tweaking two harness settings tripled those scores. However, Nvidia achieved a 100 percent score by adding a supervising agent. El Hallak noted this supervisor acts like a CEO to nudge the working agent away from dead ends.
Most current agent users rely on a single harness layer. Nvidia researchers built their own system called Agentic Variation Operators. This is not a commercial product. Nvidia produces open tech for harnesses under the Nemo brand.
Databricks CEO Ali Ghodsi told TechCrunch that harness choice also impacts AI costs. Choosing the wrong harness can double expenses. El Hallak emphasized that open harnesses give users more control over accuracy. Nvidia plans to release open bits and pieces for building harnesses.
