Nvidia researchers built an agent architecture called AVO that scored a perfect 100 on ARC-AGI-3, an interactive reasoning benchmark, using Claude Opus 5 as its underlying model, according to Nvidia.
AVO, short for Agentic Variation Operators, was tested on ARC-AGI-3's unlabeled 2D games, according to a Nvidia Developer Blog post and the accompanying research paper.
Claude Opus 5 alone scores about 30% on the same public benchmark set at high reasoning effort, according to Nvidia. Wrapped in AVO, the same model cleared all 183 levels across 25 environments in 6,624 actions, about 12% fewer actions than the previous best system, Nvidia said.
AVO adds two components around the base model: a persistent memory system that carries forward prior attempts, test results and reasoning so the agent does not repeat failed exploration, and a supervisor that monitors the agent's broader trajectory and redirects it when progress stalls. "The model matters, but the model is not the entire agent," the paper's authors wrote.
Adel El Hallak, vice president of product at Nvidia's AI unit, told TechCrunch the result reframes what counts as an AI agent: not just the model answering a prompt, but the scaffolding of tools and checks built around it.
For builders choosing where to spend engineering time, Nvidia's own benchmark says the harness closed a 70-point gap the model itself could not close on its own, on a task Nvidia designed and ran itself. That is worth reading as a single vendor's demonstration, not an independent result, but the direction it points, invest in the scaffolding rather than wait for the next model, matches what other labs have been saying all year.