Nvidia reported that a custom agent harness using Agentic Variation Operators enabled Claude Opus 5 to achieve a 100% score on the ARC-AGI-3 benchmark
Benchmark result ● Confirmed 86% confidence first seen
Nvidia published findings that a specially designed agent “harness,” including memory handling and supervisory components, dramatically improved long-horizon performance when running Claude Opus 5 on ARC-AGI-3. The reported score increased from about 30% without the harness to 100% with the harness, suggesting that agent system design can be a key driver of benchmark outcomes.
AI harnesses
Claude Opus 5
NVIDIA
agent harness
ARC-AGI-3
AVO
Agentic Variation Operators (AVO)
Claude Code
Claude Opus 4.5
Claude Sonnet 5
Harness-Bench
SWE-Bench Pro
The New Stack
Source coverage
Claude Opus 5 scored 30% on ARC-AGI-3. Wrapped in Nvidia’s AVO, it hit 100%.
The New Stack
Nvidia just showed that the harness, not the AI model, is now the real hero
TechCrunch
independent
The Evolution of the Agent Harness
Latent Space
newsletter
AI Harnesses Are the Software Shells That Turn AI Models Into Capable Agents
Trending Topics
independent