NVIDIA says its Agentic Variation Operators system, or NVIDIA AVO, achieved a perfect score on the public set of ARC-AGI-3, completing all 183 levels across 25 interactive environments. The result is significant because it frames long-horizon agent performance as a systems problem rather than a simple measure of the underlying language model.

The claim comes from a NVIDIA Developer Blog post describing AVO as a general-purpose architecture for sustained autonomous work. The company says the same system also optimized GPU kernels over a seven-day run, suggesting that its design can transfer between software engineering workflows and unfamiliar interactive reasoning tasks. However, the performance figures are NVIDIA-reported, and the available evidence does not include independent validation or a peer-reviewed assessment.

A perfect ARC-AGI-3 result from a broader agent system

NVIDIA reports that AVO recorded a 100.00 RHAE score on the ARC-AGI-3 public set. It says the agent completed every level in the benchmark’s 25 environments and used 12% fewer environment actions than VISTA, a comparison system referenced in the company’s post.

ARC-AGI-3 is designed around unfamiliar interactive environments in which agents must infer how actions work and discover what they are expected to accomplish. Unlike a conventional question-answering test, the benchmark requires an agent to maintain progress over multiple steps while learning from feedback.

NVIDIA’s central argument is that this kind of evaluation measures more than model intelligence. The agent must decide what to observe, which tools to use, how to preserve useful state, and how to recover when an approach fails. In the company’s description, the surrounding harness determines how effectively a model’s capabilities become sustained autonomous behavior.

The post also says that a system built around Claude Opus 5 raised performance from a 30% model baseline to the reported 100% result when incorporated into the complete AVO architecture. That comparison is presented by NVIDIA as evidence that system design can materially change benchmark outcomes. It should not be read as an independent ranking of either Claude Opus 5 or AVO without more detail about the evaluation setup.

How Agentic Variation Operators works

AVO is described as a coding and task-execution system that can inspect and edit code, run commands, consult documentation, and validate changes through execution. Its defining feature is not a single tool, NVIDIA says, but the ability to continue working across long sequences of experiments.

Two mechanisms receive particular emphasis: persistent memory and supervision. Persistent memory stores earlier implementations, test results, compiler and profiler output, and accumulated reasoning. This allows the agent to resume from the current state instead of rebuilding its understanding after each context window.

A supervisory component watches the broader trajectory. When the main agent becomes stuck or repeats unproductive cycles, the supervisor can redirect it toward a different strategy. The main agent remains responsible for choosing what to inspect, modify, test, and commit, while the supervisor monitors progress across the longer run.

For ARC-AGI-3, NVIDIA says it connected the same underlying agent to a different interface, changing the available environment tools and evaluation process while retaining the core architecture. That portability is the more consequential claim for AI builders: the system is intended to be adapted to new feedback loops without redesigning the entire agent from scratch.

Evidence from GPU-kernel optimization—and its limits

NVIDIA first describes AVO through a GPU-kernel optimization experiment. In that work, the system autonomously explored more than 500 optimization directions over seven days and committed 40 kernel versions, according to the company.

On NVIDIA DGX B200 systems, NVIDIA reports that the resulting multihead attention kernels performed up to 3.5% better than cuDNN and up to 10.5% better than FlashAttention-4 across the configurations it evaluated. The company also says AVO adapted the evolved kernel to grouped-query attention in about 30 additional minutes of autonomous work.

These figures illustrate the type of workflow AVO is built to handle: make a change, execute a hardware-grounded test, interpret the result, and decide what to try next. They do not establish that the system will deliver the same gains across other kernels, hardware configurations, or production codebases. The reported comparisons are vendor benchmarks, and the post does not provide enough detail in the supplied evidence to independently assess the full test methodology.

The ARC-AGI-3 result has a similar qualification. A perfect public-set score is a strong signal within that benchmark, but it does not by itself prove general-purpose reliability in enterprise environments. Real deployments introduce costs, permissions, tool failures, sensitive data, ambiguous objectives, and requirements for auditability that a public benchmark may not capture.

What the result means for builders and enterprises

For product teams, the immediate lesson is architectural. Increasing the capability of a base model may improve individual responses, but long-running workflows also need state management, controlled tool access, execution feedback, checkpointing, and recovery logic.

That matters for coding assistants, research systems, data-analysis tools, and other AI agents that cannot succeed with a single model response. An agent optimizing code needs to preserve test history and distinguish a genuine improvement from a noisy result. An enterprise workflow needs similar controls to prevent repeated actions, contain failures, and explain why a system took a particular path.

AVO’s design also highlights a deployment trade-off. Persistent experimentation can improve results, but it consumes compute and may increase latency. A seven-day autonomous optimization run is suitable for some engineering searches, but not for every customer-facing task. Builders will need policies for when an agent may continue independently, when it must ask for approval, and how much budget it can consume.

The architecture could also sharpen competition between model vendors and infrastructure providers. If agent performance increasingly depends on the harness around a model, model benchmarks alone become less predictive of product performance. Companies may compete through memory systems, evaluators, supervisors, tool orchestration, and execution environments as much as through model training.

What to watch next

The first signal to watch is independent reproduction of the ARC-AGI-3 result, including the exact model, prompts, tools, action budgets, and supervision configuration. Those details will determine how much of the gain comes from AVO’s general architecture and how much comes from task-specific engineering.

The second is evidence outside NVIDIA-controlled demonstrations. Useful follow-up data would include performance on additional interactive benchmarks, failure rates over long runs, compute consumed per completed task, and comparisons against other agent harnesses.

For enterprise buyers, deployment evidence will matter more than a single benchmark score. Watch for information about permission controls, audit logs, sandboxing, recovery from tool errors, data isolation, and the frequency with which human intervention is required.

Finally, the GPU-kernel work provides a practical test of whether autonomous optimization can produce repeatable gains across different workloads and hardware. Results beyond the reported configurations would make the engineering claim more durable.

Creati.ai perspective

NVIDIA’s announcement is best understood as a report about agent architecture, not proof that one model has suddenly become universally capable. The reported ARC-AGI-3 score is notable because AVO combines memory, supervision, tools, and feedback into a system designed to keep working after the first answer fails.

For AI builders, the durable takeaway is that long-horizon reliability may depend on the surrounding control loop as much as on the model. NVIDIA has supplied an interesting vendor benchmark; the next question is whether independent teams can reproduce the result economically, safely, and across tasks that matter outside the benchmark.