Nvidia's AVO harness reportedly completes public ARC-AGI-3 environments

Forbes reports that Nvidia's AVO agent harness, running Anthropic's Claude Opus 5, completed every level in ARC-AGI-3's 25 public environments in 6,624 actions. The article says the underlying model was not retrained or fine-tuned; AVO instead supplies persistent memory, supervision, an iterative inspect-plan-act-evaluate loop and swappable tools. It reports that the result used about 12 percent fewer actions than the prior most efficient system, while the bare model completed about 30 percent of the benchmark. Nvidia initially built AVO to optimize GPU code, the article says.