NVIDIA AVO Agent Scores 100% on ARC-AGI-3 Test

NVIDIA's AVO coding agent has achieved a perfect 100% score on a public test dataset called ARC-AGI-3. What's interesting is it completed all 183 levels of 25 games without any prior instructions. Here's the thing: AVO is a "harness" built around Anthropic's Claude Opus 5 model. The Opus 5 model alone scored only 30% on the same test. A harness is like an external layer of software that helps an AI model connect to the outside world, create self-correcting loops, manage memory, and structure prompts. NVIDIA originally designed AVO to optimize CUDA GPU kernels. In that role, it ran autonomously for seven days, explored over 500 directions, and produced kernels that outperformed FlashAttention-4 by up to 10.5%. Without changing its core design, AVO was then applied to the ARC-AGI-3 visual and interactive logic puzzles. It successfully transferred its code-inspecting and self-correcting abilities. The AVO solved these 183 levels using 6,624 actions, which is a 12% increase in efficiency compared to other leading agent wrappers. The bottom line: This shows the significant value an effective AI harness can add to an existing AI model's capabilities.