Agentic Variation Operators, or AVO, started as a tool for a narrow, technical job: making GPU kernels faster. NVIDIA built it to inspect existing kernel code, propose a change, run the change against real hardware, and use what it learns to decide the next attempt, repeating that loop for as long as the task requires. The company then aimed the same architecture at ARC-AGI-3, a benchmark where Claude Opus 5 manages about 30 percent of the levels running on its own.

An Agent Built to Replace a Fixed Search Step

Evolutionary search has solved hard engineering problems for decades. Generate a batch of candidate solutions, test each one, keep the winners, and vary them again. The step that generates new candidates has followed fixed rules for decades: mutate this parameter, cross these two solutions, repeat. AVO replaces that fixed step with an agent that decides for itself what to inspect, what to change, what to test, and what to keep, based on what it learned from the attempt before.

Two mechanisms keep the decision-making stable across long runs. Persistent memory carries forward prior implementations, test results, and profiler output, so the agent resumes from where it left off instead of rebuilding its understanding each time. A supervisor watches the broader search for stalled progress and redirects the main agent toward a different approach when the current one stops producing gains.

NVIDIA put both mechanisms to a real test in the GPU-kernel work. Across a seven-day run with no interruption on NVIDIA's DGX B200 systems, AVO explored more than 500 optimization directions and committed 40 kernel versions. The resulting attention kernels outperformed cuDNN by up to 3.5 percent and beat FlashAttention-4 by up to 10.5 percent. The agent then adapted its own evolved kernel to a related attention variant in about 30 minutes of additional autonomous work.

Applied to a Benchmark It Was Never Built For

ARC-AGI-3, built by researcher François Chollet's ARC Prize Foundation, drops an agent into a game environment with no manual. No stated rules, no listed objective, no description of what the available actions do. The agent has to press buttons, watch what changes, and build a working theory of the world through interaction alone, across 25 separate environments and 183 levels. NVIDIA changed the task interface for this test and left the rest alone. The model reads a 64-by-64 text grid instead of source code and compiler output. The inspect-plan-implement-evaluate loop, the memory, and the supervisor carried over from the kernel work unmodified.

Scoring runs on Relative Human Action Efficiency, or RHAE, a metric that weighs whether the agent finished each level against how many actions a first-time human needed for the same level. Running Anthropic's Claude Opus 5 as its underlying model, AVO finished all 183 levels in 6,624 actions and posted a perfect 100.00 RHAE score. ARC Prize reports Claude Opus 5 scoring about 30 percent on the same benchmark at its high reasoning setting, without the AVO architecture around it.

What the Two Numbers Do and Don't Show

NVIDIA is explicit that the 30-to-100 gap is not a controlled measurement of what AVO itself contributed. The standalone score came from ARC Prize's own evaluation harness and reasoning setting. The AVO score came from a different reasoning configuration and a different agent system built around the model.

Both numbers are real, and neither one alone isolates the harness as the cause of the gap.

The 100.00 also covers the ARC-AGI-3 public set, the levels available for anyone to study and build against. NVIDIA has not run AVO against the semi-private or private competition sets, which exist to test agents on puzzles they have not been able to prepare for in advance. A separate comparison against VISTA, another Claude Opus 5 harness that completed the same public levels in 7,542 actions against AVO's 6,624, points to persistent memory as one plausible source of the efficiency gain, though NVIDIA notes the two systems differ enough in design that the comparison is not an apples-to-apples ablation either.

Why This Belongs on a CIO's Radar

NVIDIA does not sell AVO as a product. It is a research architecture, and the transfer from kernel optimization to an unrelated reasoning benchmark is the result NVIDIA's team emphasizes most: the same memory-and-supervisor machinery held up across two domains with the task interface as the one thing that changed. That is a genuine systems finding, independent of any single benchmark score.

AVO fits a broader pattern in how agent capability is built and sold. shashi.co has covered Databricks shipping Omnigent, a meta-harness designed to sit above whichever coding agent a developer already runs, and NVIDIA's own Open Secure AI Alliance work on governing agents at the identity and permissions layer. AVO adds a data point to that pattern: the harness around a model is becoming a place where real engineering work happens, separate from which model sits underneath it. For CIOs and CTOs evaluating coding agent vendors, that means the model question and the harness question are two different questions, each needing its own answer.