Nvidia just showed that the harness, not the AI model, is now the real hero
By Pradhyuman,
Nvidia study focuses on AI harnesses
Nvidia researchers published a study on the software wrapper, or harness, that controls an AI model. The researchers used their custom harness, Agentic Variation Operators, to run Claude Opus 5. The model achieved a 100% score on the ARC-AGI-3 benchmark. Without the harness, Claude Opus 5 scored 30%. The custom harness uses a supervisor component to guide the agent when it gets stuck.
Adel El Hallak of Nvidia said that an agent consists of the model and its harness.
Other organizations also reported findings about harnesses. OpenAI researchers tripled their model scores on the same benchmark by changing two harness settings. Databricks published research that showed how harness choice changes computing costs.
Maintained by Pradhyuman.
Filed under: Nvidia, Anthropic, Models & Research, Products & Agents