On August 21, NVIDIA published a technical blog post about AVO (Agentic Variation Operators), a framework that puts Anthropic's Claude Opus 5 on a tight leash—lifting the same model's ARC-AGI-3 score from a raw 30% to a perfect 100%, clearing all 25 environments and 183 levels in just 6,624 steps.
The key point? Not a single line of the model changed. What changed was the "scaffolding" around it.
A Feynman-Style Breakdown: What Is the Harness?
Think of a frontier LLM as a fine horse. It runs fast and knows the roads, but with no one holding the reins, it will bolt across the plains, run into dead ends, or circle back to where it started.
What is ARC-AGI-3? A collection of 2D mini-games with no instruction manuals. The model enters, sees neither rules nor goals, and must infer the environment's dynamics and win conditions purely by trial and error. This tests exactly "long-horizon tasks"—chaining hundreds or thousands of decisions without losing the thread.
Claude Opus 5 running raw scored 30%—already the highest among all tested models. OpenAI's models scored below 10%, which frustrated them enough that they published research last month showing that adjusting just two harness settings could triple their scores.
NVIDIA's AVO does two things to make the harness actually work:
One is persistent memory. Ordinary agents rebuild context from scratch every round, like an amnesiac re-reading their own notes each day. AVO carries "previous implementations, evaluation results, and reasoning chains" forward continuously, picking up where it left off instead of starting over.
The other is a supervisor—the key to the perfect score. It acts like a CEO: it does no work itself, only watches the main agent's "search trajectory." When the main agent gets stuck, starts circling, or heads into a dead end, the supervisor nudges it back onto an effective path. In the seven-day GPU kernel optimization experiment, the main agent decided on its own what to modify, test, and submit, while the supervisor only stepped in when progress stalled. This division of labor let the loop run for over a hundred rounds without derailing.
On GPU kernel optimization, AVO ran continuously for seven days, explored over 500 directions, submitted 40 kernel versions, and achieved up to a 10.5% speedup over FlashAttention-4 on a DGX B200. Plugged into the ARC-AGI-3 interface, the same architecture scored a perfect 100%.
Why This Matters More Than "Another Benchmark Win"
Models are becoming commodities. Everyone competes on parameters and pretraining data, but NVIDIA's numbers deliver a pointed message: the bottleneck for long-horizon tasks isn't the brain—it's the cage.
- The same Claude Opus 5, with or without the harness, differs by 70 percentage points.
- Databricks previously showed that with the same model, choosing the wrong harness inflates costs substantially.
- In April, Microsoft tested 19 models on long-horizon document editing—including frontier models—and all of them corrupted the documents. Not because they weren't smart, but because no one was managing their attention, memory, and direction.
- AVO is not an NVIDIA product—it's a research artifact. What NVIDIA actually sells are the harness-building components under the NeMo brand, many of them open source. But the directional signal is clear.
- This adds a footnote to today's AI coding boom: "autonomous delivery" tools like oncall-kit and Copilot Autopilot have moats not in models but in the orchestration layer. Whoever's harness lets agents run for days without crashing, circling, or deleting the database wins.
- The ARC Prize team is probably already revising the benchmark. A benchmark designed by humans to stump AI designed by humans was always an experiment destined to expire.
This means the center of the "architecture war" is shifting: from "breeding better horses" to "building better cages." Memory management, supervisor architectures, and failure recovery will attract the kind of investment and competitive pressure once reserved for the models themselves.