A custom harness with memory and a supervisor component took Claude Opus 5 from 30% to 100% on the interactive reasoning benchmark.

Nvidia researchers published findings showing that a custom harness with a supervisor component lifted Claude Opus 5 to a perfect 100% score on the ARC-AGI-3 interactive reasoning benchmark, up from a 30% score without the harness.

Adel El Hallack, vice president of product in Nvidia’s AI unit, said an agent is the model plus the scaffolding around it, which the company calls the harness, including tools, runtime, skills, and libraries.

For builders and operators, the takeaway is that agent performance depends heavily on harness design, not just the underlying model, especially for long-horizon tasks requiring memory and feedback management.

Watch for more agent frameworks to adopt supervisor components and memory optimizations in the coming months, as benchmarks like ARC-AGI-3 expose the gap between raw model scores and agent system results.

What matters

• Nvidia’s custom harness with a supervisor component lifted Claude Opus 5 to a perfect 100% score on ARC-AGI-3.
• For builders, the harness, not the model, governs success on long-horizon agent tasks.
• Watch for more agent frameworks adding supervisor components that keep models on task.

Why it matters

Watch for more agent frameworks adding supervisor components that keep models on task.

This GenAI News article was prepared in original wording using reporting and materials published by TechCrunch AI. Source reference: https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/.

Drafted by the GenAI News review pipeline.