Nvidia Harness Lifts Claude Opus 5 to 100% on ARC-AGI-3
Nvidia researchers say a custom agent harness helped Claude Opus 5 score 100% on ARC-AGI-3, an interactive reasoning benchmark made up of 2D games without instructions. The same model scored 30% without the harness, which Nvidia described as the best result among the models tested without that additional system layer.
The work centres on Nvidia’s Agentic Variation Operators, or AVO, a research harness rather than a new Nvidia product. Its result supports a growing argument in agentic AI: the model matters, but the surrounding system can have an outsized effect when an agent must make many connected decisions over an extended task.
What Nvidia changed around the model
Adel El Hallack, vice president of product in Nvidia’s AI unit, described an agent as more than a model endpoint. In his account, the full system includes the model, the scaffolding or harness, the tools it can use, its runtime, and the skills and libraries available to it.
For AVO, a key addition was a supervising agent alongside the main agent. The supervisor can nudge the working agent when it is stuck, heading towards a dead end, or revisiting a path it has already explored. Nvidia argues that this kind of orchestration is particularly relevant for long-horizon work, where a system has to sustain decisions rather than simply return one response to a prompt.
ARC-AGI-3 offers a demanding setting for that claim. A perfect score means the system can play and win the benchmark’s games at a human level. OpenAI had reported that adjusting two harness settings tripled its models’ scores, after results below 10%, but Nvidia’s reported 100% result went further.
Why harness design is becoming a business issue
The findings arrive as enterprises assess AI agents for tasks involving document work, tools and multi-step execution. Microsoft research published in April tested 19 large language models on long-horizon document-editing tasks and found errors in the resulting documents, including from frontier models. The central challenge is not only generating a capable answer, but keeping an agent on track over a chain of actions.
The competitive pressure around enterprise agent systems is also visible in Microsoft’s enterprise AI competition, where Microsoft’s position alongside OpenAI and Anthropic illustrates how model ecosystems and deployment choices are converging in business technology.
Nvidia also pointed to cost implications. Databricks CEO Ali Ghodsi said that different harnesses used with the same model can materially change cost, potentially by two times. Nvidia produces open building blocks for harness construction under the NeMo brand, with some commercial and much openly available technology.
Control beyond model selection
Nvidia’s broader position is that open harnesses give users more controls over accuracy, infrastructure and runtime behaviour. That does not remove the importance of selecting a model, but it shifts evaluation towards the complete agent stack and the safeguards governing it.
For businesses, the practical implication is to test agent systems as integrated deployments: assess the model alongside memory, tool access, supervisory logic, runtime controls and cost behaviour before assigning long-horizon work.