NVIDIA Scores 100% on ARC-AGI-3, and the Hero Is the Harness, Not the Model

NVIDIA says its agentic architecture AVO has reached 100% on the ARC-AGI-3 benchmark, calling it a demonstration of a "frontier-level general-purpose architecture for long agentic tasks." The company's engineers laid out the details on the NVIDIA Developer blog.

What happened

ARC-AGI-3 is the third generation of François Chollet's abstract-reasoning benchmark and one of the most stubborn tests in the industry. Unlike its predecessors, it consists of interactive game environments in which the agent has to work out the rules of the world on its own by experimenting with it, with no instructions or hints. Frontier models have long posted modest results here: until recently, scores below 40% counted as a notable achievement.

Against that backdrop, AVO's 100% is a sharp break. But the key part of the announcement is not the number. It is the explanation of how the number was achieved.

• ARC-AGI-3 is an interactive benchmark: the agent infers the environment's rules on its own, by trial and error
• AVO posted a perfect score of 100%
• According to NVIDIA, the decisive contribution came from the agent's architecture, not from the model's "raw" power

The harness is the new center of gravity

NVIDIA's central thesis: a frontier language model is just one component of an AI agent. Everything else is decided by the surrounding system, which the industry calls the harness. It determines how the model receives context, how it calls tools, how it preserves state between steps, how it reacts to feedback from the environment, how it recovers from failures, and how it holds on to progress in tasks that run for dozens or hundreds of steps.

TechCrunch put the conclusion even more bluntly: NVIDIA has shown that the real hero now is the harness, not the model. The same "brain" in a different harness produces radically different results. That helps explain why companies building agentic products increasingly compete on engineering of the surrounding loop rather than on access to the latest model.

What this means

AVO's result shifts the focus of the race from training models to the systems engineering around them. That is a discipline in which teams with a strong infrastructure culture gain the advantage, not just the labs with the largest clusters. There is a flip side, too: once a benchmark is maxed out at 100%, it stops measuring progress. ARC-AGI-3 now faces the fate of its predecessors, and the industry again needs a test that nobody has solved yet.

Frequently asked questions

What is ARC-AGI-3?

An interactive general-intelligence benchmark from François Chollet's ARC series: a set of game environments whose rules the agent must work out on its own by interacting with the world. It is considered one of the hardest tests of generalization rather than memorized knowledge.

What is a harness?

The software scaffolding around an AI agent: the subsystems that feed it context, call tools, handle memory, process errors and supervise long tasks. NVIDIA's result shows that the quality of the harness affects the outcome no less than the quality of the model itself.