Nvidia's harness pushes Claude Opus 5 from 30% to 100% on ARC-AGI-3
Nvidia researchers wrapped Claude Opus 5 in a custom harness called Agentic Variation Operators — better memory handling plus a 'supervisor' agent that nudges the main agent off dead ends — and took its ARC-AGI-3 score from 30% to 100%. OpenAI's own tuning of two harness settings tripled its models' scores last month but never approached 100%, so the harness layer, not the base model, is where long-horizon agent performance is currently won.
Source: techcrunch.com ↗
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.
Adel El Hallak, VP of Product, Nvidia AI
Why this matters
- → Harness design, not base model, unlocked perfect ARC-AGI-3 score
- → Supervisor agent component prevents long-horizon task drift
- → OpenAI's tuning approach tripled scores but never reached parity
Scaffolding beats brains