Nvidia's custom-built harness took Claude Opus 5 from 30% to a perfect 100% on the interactive reasoning benchmark ARC-AGI-3, with nothing changed in the model itself.
The 30% run was already the strongest result of any model Nvidia's researchers put through the trials. They named that harness the Agentic Variation Operators, or AVO. Their research, published last week, argues that on long-horizon work — tasks that chain many decisions together, sometimes across days, before any finished work comes out — the choice of harness weighs heavier than the choice of model.
ARC-AGI-3 is built from a collection of 2D games that arrive with no explanation of the rules, leaving the model to work out for itself how they are played and how to win. A 100% score means it wins them as capably as a person can.
The harness is the scaffolding that handles memory, context and feedback, and it is what turns a model into an agent at all. Adel El Hallack, a vice president of product at Nvidia's AI unit, told TechCrunch that the world by and large treats an agent as little more than an API call to the model, when an agent takes in considerably more: the model, the tools built around it that Nvidia calls the harness, the runtime, and the skills and libraries it is handed access to.
A harness also has to watch the agent it runs. El Hallack said the more interesting piece of the work was setting a supervising agent alongside the main agent doing the job, in the role of a chief executive: it nudges the agent when it strays off direction, sets off down a route that may end up going nowhere, or goes back over ground it has already covered. The concept is not new, but most agent users today rely on a single layer for their harness, such as Claude Code, Codex or Hermes.
Nvidia's result adds to a swelling body of evidence that the model you pick is nowhere near the only thing determining how an agent performs. OpenAI, whose own models came in under 10% on ARC-AGI-3, ran a study of the benchmark last month and found that tweaking two settings on the harness tripled those scores — though not one model in that study got anywhere near a perfect 100%. In July, Databricks published research indicating that the harness, far more than the model, has a dramatic effect on what AI costs.
The harness that closed the gap is not on offer, though. AVO is a research build rather than a fresh Nvidia product, and what the company puts out instead is an assortment of open technology for assembling harnesses under the Nemo name, some of it sold commercially and a large share of it free.