Nvidia cracked a reasoning benchmark. A software harness, not a new model, made the difference

From 30 to 100% with a software harness

Claude Opus 5 on its own solved about 30% of the complex logic tasks in ARC-AGI-3, a set of 2D games without instructions. Nvidia researchers then placed the model inside a custom memory-management system and added a supervising agent. The code did not give the model new knowledge, but it helped the model stay on course and leave dead ends.

With that external control, the same model reached 100%. The result does not mean model choice is irrelevant. It does show that, on long-horizon tasks, the surrounding architecture can change performance as dramatically as the model weights themselves.

The model is only one layer of an agent

Nvidia's work highlights the broader composition of an agentic system. The model supplies reasoning capability, while the software harness manages memory, context, tools, and feedback. That layer turns a response to one prompt into a process that can connect many decisions and recover from wrong turns.

The supervising agent was central to the result. When the main agent became stuck or retraced a failed path, another layer nudged it elsewhere. Performance therefore emerged from the interaction of several system components, not from the model alone.

Architecture changes both performance and cost

A software harness is not a free layer. Managing context, tools, and feedback adds operations, and implementation choices can materially change runtime cost. TechCrunch cites Databricks research suggesting that the same model can cost twice as much with the wrong harness.

The price of a model call therefore does not capture the economics of an agent by itself. Teams also need to track step counts, repeated attempts, and supervisory overhead, because this is where the gap between two architectures can multiply.

Open harnesses become a competitive layer

Nvidia did not present the result as a new finished product. It built a research system called Agentic Variation Operators and offers both open and commercial components for similar harnesses under its NeMo brand. Its broader argument is that an open agent stack gives users control over the harness, infrastructure, and runtime.

The next shift will therefore happen beyond model leaderboards. Comparisons of performance, cost, and safety will also include how different agent stacks handle memory, supervision, and tools.

Lilith's verdict

Claude Opus 5 scored 30% without the harness and 100% with supervision. Humanity has rediscovered that even a digital genius performs better when someone stops it from wandering off.