NVIDIA researchers have demonstrated that the architecture surrounding a frontier language model can matter as much as the model itself. On August 21, the company’s Agentic Variation Operators system, known as AVO, recorded a perfect 100.00 Relative Human Action Efficiency score on the public set of the ARC-AGI-3 benchmark.

AVO completed every one of the 183 levels across all 25 interactive environments. The underlying model was Anthropic’s Claude Opus 5, the same model that scores approximately 30.2 percent when evaluated on its own under high-reasoning settings. The gap between those two results was closed entirely by the agent system NVIDIA wrapped around the model.

ARC-AGI-3, developed by the ARC Prize Foundation led by François Chollet, is deliberately designed to resist the usual benchmark inflation techniques. Agents are dropped into unfamiliar, abstract, turn-based environments with no instructions, no stated goals, and no prior knowledge of the rules. They must explore, infer dynamics, discover objectives, and act efficiently. Humans solve 100 percent of the environments. Until recently, the best standalone frontier models remained in the single digits or low teens.

Claude Opus 5 had already set a new model record at 30.2 percent. NVIDIA’s result shows what happens when that same model is placed inside a carefully engineered long-horizon agent loop.

The AVO architecture centers on an iterative cycle of inspecting context, planning, implementing changes, and evaluating outcomes. Persistent memory carries forward prior results, implementations, and reasoning so the agent does not have to reconstruct its understanding from scratch at every step. A supervisor monitors the broader trajectory for stagnation or repeated unproductive cycles and can redirect the main agent when progress stalls.

NVIDIA first validated this design on a demanding real-world engineering task: autonomous GPU-kernel optimization. Over seven continuous days, AVO explored more than 500 optimization directions and committed 40 kernel versions. On NVIDIA DGX B200 systems, the resulting multi-head attention kernels outperformed cuDNN by up to 3.5 percent and FlashAttention-4 by up to 10.5 percent. The same system later adapted the evolved kernel to grouped-query attention in roughly 30 minutes of additional autonomous work.

That engineering loop transferred directly to ARC-AGI-3. The team kept the core agent unchanged and only swapped the environment-specific tools and evaluation interface. Environments were presented as text-only 64-by-64 grids. AVO finished the full public set in 6,624 environment actions. By comparison, the previous strong agent harness known as VISTA required 7,542 actions with the same Claude Opus 5 backbone—an efficiency gain of about 12 percent.

NVIDIA is careful to note the limitations. The 100 percent score applies only to the public set. No private or hidden competition set was submitted. The comparison with VISTA is not a controlled ablation; the two systems differ in observation representation, memory management, and other components. The result is therefore best read as evidence that system design can unlock large gains in sustained autonomous performance, not as a claim that ARC-AGI-3 has been solved in the absolute sense.

The broader industry implication is clear. As models become more capable, the bottleneck is shifting from raw intelligence to the scaffolding that turns that intelligence into reliable, long-horizon action. Coding agents, research agents, and enterprise automation systems all face the same core requirements: maintaining state across many steps, recovering from mistakes, and continuing to make progress when the path is not obvious.

AVO is one concrete demonstration that a general-purpose agent architecture, originally built for software engineering and hardware optimization, can transfer to abstract interactive reasoning without domain-specific redesign. That transferability is the part of the result that is likely to influence how other labs and startups design their own agent stacks in the months ahead.

For practitioners building agentic systems today, the takeaway is practical. Investing in persistent memory, supervisory oversight, and a robust inspect-plan-implement-evaluate loop can produce performance jumps that rival or exceed incremental model upgrades. The model remains essential. The system around it increasingly determines whether that model’s capability is actually usable on hard, multi-step work.

NVIDIA has published the technical details and architecture description in its developer blog and related research materials. The ARC-AGI-3 public set remains available for independent evaluation, ensuring the claim can be examined and challenged by the wider research community.