NVIDIA’s AVO Coding Agent Scores 100% on ARC-AGI-3 Public Benchmark, Showing the Power of AI Harnesses
NVIDIA has delivered a striking example of how much an AI system can improve when a strong software harness is built around it. The company’s AVO coding agent was tested on the ARC-AGI-3 public dataset and successfully completed all 183 levels across 25 publicly available games with a perfect 100 percent score.
What makes the result especially notable is that AVO reportedly received no prior instructions, rules, or stated goals. Instead, it had to learn by experimenting, observing what happened, remembering useful information, and correcting its mistakes along the way.
AVO is built around Anthropic’s Claude Opus 5 model. On its own, Claude Opus 5 scored around 30 percent on the same ARC-AGI-3 public set. With NVIDIA’s AVO harness wrapped around it, the system reached a flawless score. That gap highlights an increasingly important point in artificial intelligence: the underlying model matters, but the system built around the model can be just as critical.
An AI harness is essentially an external software layer that gives a model tools, structure, memory, and feedback loops. If the AI model is the brain, the harness acts like the body, workspace, toolkit, and operating strategy that allow it to interact with complex tasks in a more useful way.
A standard language model cannot independently browse an interface, run scripts, inspect files, test solutions, or keep organized records of long problem-solving sessions. A harness can provide those abilities. It connects the model to an external environment, lets it take actions, gathers the results, and feeds the information back into the model so it can improve its next attempt.
This is especially important for tasks that require trial and error. Instead of producing a single answer and stopping, a harnessed agent can test an idea, analyze why it failed, adjust its approach, and try again. Over time, this creates a self-correcting loop that can turn a general-purpose AI model into a more capable problem solver.
Memory management is another major advantage. As AI models process more information, they can lose track of earlier details or become distracted by irrelevant errors. A well-designed harness can work like a structured notebook, saving useful discoveries, filtering out noise, keeping track of successful strategies, and helping the model maintain a clear path toward solving the task.
NVIDIA originally developed AVO for a very different purpose: optimizing CUDA GPU kernels. In that setting, the agent ran autonomously for seven days, explored more than 500 possible optimization paths, and produced GPU kernels that reportedly outperformed FlashAttention-4 by as much as 10.5 percent.
For the ARC-AGI-3 test, NVIDIA did not redesign the core agent architecture. Instead, it replaced the GPU engineering tools with the ARC-AGI-3 task interface. That allowed AVO to apply its existing strengths, including code inspection, experimentation, memory, and self-correction, to a completely different type of challenge: interactive visual logic puzzles.
The result was a perfect run through the public ARC-AGI-3 dataset. AVO solved all 183 public levels without being told the rules or objectives in advance. It learned through action, feedback, and adaptation.
NVIDIA also reported that AVO completed the 183 levels using 6,624 actions. That makes it more efficient than some other leading agent wrappers, including VISTA, which required 7,542 actions to complete the same public set. In practical terms, AVO not only solved every task but did so with fewer steps.
Still, there is an important limitation. The ARC-AGI-3 platform currently does not allow custom external agent harnesses to run on its hidden private dataset. Because of that, AVO’s performance on the more challenging private evaluation set remains unknown.
Even with that caveat, the public result is significant. It suggests that future progress in AI may not come only from building larger models, but also from designing better systems around them. A powerful model can generate ideas, but a strong harness can help it test those ideas, remember what works, recover from errors, and keep improving until the problem is solved.
NVIDIA’s AVO demonstration shows why AI agents are becoming a major focus across the industry. The next leap in artificial intelligence may depend not just on smarter models, but on smarter ways to guide, equip, and organize them.
