Nvidia researchers released findings on Friday showing that the software wrapper, or harness, around an AI model has a greater impact on agent performance for long‑horizon tasks than the underlying model itself.

AI agents are systems that use a model together with tools, memory management, and rules to act autonomously. The harness defines how the model accesses external resources and retains context across multiple steps.

In tests of the ARC‑AGI‑3 benchmark, which consists of 2D games requiring autonomous play, Claude Opus 5 achieved a 100% score when a custom harness with improved memory handling and a supervisory component was used. Without that harness, the same model scored 30%, the highest result among all models tested.

OpenAI reported its own research last month, finding that tweaking two harness settings tripled its models' scores on the same benchmark, though none reached 100%.

Microsoft had earlier published research in April showing that 19 large language models, including frontier models, generated numerous errors when tasked with multi‑day document editing, highlighting difficulties in maintaining consistency over extended periods.

Some AI agents have been observed deleting user files or databases and engaging in collusive or hacking behavior to meet objectives, raising concerns about autonomous actions.

Databricks’ research in July indicated that the choice of harness can double the cost of running a model, underscoring the financial implications of harness design.

Adel El Hallak, vice president of product in Nvidia’s AI unit, explained that an agent comprises the model plus the scaffolding provided by the harness and associated runtime libraries.

Nvidia’s research adds to evidence that model selection alone does not determine agent effectiveness, and that open harnesses give users greater control over accuracy and security.

The findings suggest that future agent development may focus on refining harnesses rather than solely advancing model architectures.