NVIDIA's AVO harness turns a 30% Claude Opus 5 into a 100% ARC-AGI-3 solver
Same model, different scaffold: the agent-architecture layer is now the load-bearing variable for long-horizon tasks.
NVIDIA reported Friday that its AVO agent system scored a perfect 100.00 RHAE on ARC-AGI-3, clearing all 183 levels across 25 environments using Claude Opus 5 as the backbone. The same Claude Opus 5, run standalone, scores roughly 30% on the benchmark (30.2% at high reasoning effort, per aiweekly.co’s readout). The 70-point delta didn’t come from the model.
It came from the scaffold around it.
AVO (Agentic Variation Operators) wasn’t built for reasoning puzzles. It started as a GPU-kernel optimization agent for Blackwell hardware, where NVIDIA reports it ran for seven days, explored more than 500 optimization directions, and produced 40 committed kernel versions that beat FlashAttention-4 by up to 10.5%. Applying it to ARC-AGI-3 required swapping environment-specific tools. The core loop, a main agent that inspects, plans, implements, and evaluates, wrapped in persistent memory and a supervisor process that intervenes when progress stalls, stayed intact.
NVIDIA credits the supervisor. “It almost acts like a CEO to nudge the agent when it goes off direction, or re-explore a path that it had previously trod,” El Hallak of NVIDIA told TechCrunch. AVO also completed the public set in 6,624 environment actions versus VISTA’s 7,542, a roughly 12% efficiency gain on the same underlying model.
The result lands inside a broader repositioning. OpenAI’s models scored under 10% on ARC-AGI-3 before harness tweaks tripled their results, though none hit 100%. Databricks reached a parallel conclusion in July, framing harness choice as a primary cost lever. NVIDIA’s own companion post argues that scaffold design (memory, supervision, action rendering) accounts for double-digit swings on identical base models.
AVO isn’t a product. Harness components ship under the Nemo brand, and the accompanying arXiv paper reads as reference architecture. That’s the point. If the model is a commodity input, the agent layer is where margin, differentiation, and vendor lock-in relocate, and NVIDIA would prefer that layer live on its stack.
Sources
- https://developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/
- https://arxiv.org/abs/2603.24517
- https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/
- https://aiweekly.co/alerts/nvidias-avo-hits-100-on-arc-agi-3-uses-12-fewer-actions
- https://developer.nvidia.com/blog/six-agent-harness-capabilities-for-higher-model-performance/