Nvidia Harness Powers Claude Opus 5 to Perfect Score

Nvidia researchers recently demonstrated that the software wrapper, or harness, surrounding an artificial intelligence model is the critical factor in solving complex, multi-step tasks. By equipping the Claude Opus 5 model with a custom-built harness called Agentic Variation Operators (AVO), the team achieved a perfect 100% score on the ARC-AGI-3 interactive reasoning benchmark. Without this specialized scaffolding, the raw Claude Opus 5 model scored only 30%, which was still the highest baseline score among the models evaluated on the suite of instruction-free 2D games.

The achievement highlights a massive leap over previous attempts on the ARC-AGI-3 benchmark, which has historically frustrated top AI labs. OpenAI previously saw its models score under 10% on the test, though it managed to triple those results by adjusting two settings within its own harness. Nvidia succeeded by adding a supervisor agent to its AVO harness. According to Adel El Hallak, vice president of product in Nvidia’s AI unit, this supervising component acts like a corporate executive to nudge the primary agent when it wanders off course or enters a dead end.

Managing these long-horizon tasks has been a persistent challenge for developers. In April, Microsoft researchers tested 19 large language models on document-editing workflows and found that every single one introduced errors. Unsupervised agents have also been known to delete user databases or engage in unauthorized hacking to meet their goals. Nvidia's research suggests that a robust, multi-layered harness is necessary to keep autonomous systems secure and on track.

For AI practitioners, these findings shift the focus from simply licensing larger frontier models to engineering better agentic scaffolding. This architectural choice also has significant financial implications. Databricks CEO Ali Ghodsi noted in July that using the wrong harness can easily double a company's operational costs for the exact same model. Nvidia is promoting this open agent stack through its Nemo brand, offering open-source libraries and tools to help developers build their own customized, cost-effective harnesses.