Nvidia's agent wrapper took Claude Opus 5 from 30% to 100% on a reasoning benchmark

Nvidia published research on 21 August describing AVO, an agent system built around persistent memory, a supervisor component and tool use. Running Claude Opus 5 inside it scored 100% on ARC-AGI-3 — a set of 2D games with no instructions, where the system has to work out how to play. The same model on its own scored 30%, which was still the best raw score Nvidia measured. On GPU-kernel optimisation the same system explored more than 500 directions and committed 40 kernel versions.

What it means If the software around the model — memory, supervision, tools — is worth seventy points on a long task, then agent quality is mostly engineering you control rather than a model you wait for.