AVO Reaches 100% on ARC-AGI-3 Public Set

NVIDIA’s AVO agent system reached a 100.00 Relative Human Action Efficiency score on the public ARC-AGI-3 benchmark, while its underlying Claude Opus 5 model scored 30.16% on its own. The result, published on August 21, 2026, puts the agent’s orchestration layer at the center of the performance story.

Benchmark Scope

AVO completed all 183 levels across ARC-AGI-3’s 25 public environments. The environments provide no prior instructions, explicit rules, or stated objectives, requiring the agent to infer how each game-like system works through interaction.

The system solved the public set in 6,624 environment actions. NVIDIA’s comparison with VISTA, which used Claude Opus 5 and required 7,542 actions, gives AVO an approximately 12% action-efficiency advantage on the same evaluation.

System	Underlying model	Public-set score	Actions
Claude Opus 5 alone	Claude Opus 5	30.16%	Not comparable
VISTA	Claude Opus 5	Not specified	7,542
AVO	Claude Opus 5	100.00%	6,624

The comparison isolates an important engineering variable. AVO did not require an updated foundation model to produce the 70-percentage-point improvement.

The Harness Architecture

AVO began as a NVIDIA system for autonomous CUDA kernel optimization. Its design uses coding agents as evolutionary “variation operators,” allowing candidate implementations to be generated, tested, retained, and improved over long search horizons.

Four capabilities explain why the system transfers to symbolic spatial reasoning:

• Persistent memory preserves code implementations, state histories, and test results across extended tasks.
• Supervisory oversight monitors search trajectories and redirects the primary agent when it stalls or reaches a dead end.
• Deterministic context and state management controls what the model sees and how feedback loops are rendered, reducing drift and wasted actions.
• Model agnosticism allows the same harness architecture to work with different frontier models.

This is the practical meaning of context engineering for autonomous systems. The model supplies proposals and local reasoning, while the harness determines how experience is recorded, when strategies change, and which state returns to the next model call.

Model Choice and Runtime Tradeoffs

NVIDIA also paired AVO with OpenAI’s GPT-5.6 Sol on challenging game subsets. Sol matched level completions faster in wall-clock time, showing that a model can matter even when the harness supplies the main structural advantage.

For production agents, the relevant comparison therefore includes more than benchmark accuracy. Developers need to measure action count, wall-clock time, model cost, recovery frequency, and the amount of supervisory computation. A system that solves a task with fewer actions may still be slower or more expensive if its oversight loop is heavy.

The pattern resembles agent evaluation, where the tested unit is the complete runtime rather than an isolated completion. It also makes agent memory an execution concern, not merely a convenience for conversation history.

Public Evaluation Boundary

The 100% result covers ARC-AGI-3’s 25-environment public set. The private evaluation restricts custom task-specific harnesses, which changes how directly the public result can be generalized to unseen benchmark tasks.

AVO’s significance comes from its claimed general-purpose architecture. The same evolutionary approach moved from continuous CUDA kernel optimization to unfamiliar spatial environments without domain-specific task code modifications, but private-set performance remains the more important test for broad generalization.

If you are building a long-running coding or research agent, benchmark the harness as a first-class system: add durable state, explicit stall detection, deterministic context assembly, and recovery policies, then measure complete task cost against a direct model baseline.