Nvidia coding agent AVO reportedly achieves 100% completion on ARC-AGI-3 public set

Nvidia’s AVO agent system reportedly solved every task in the ARC-AGI-3 public set, a benchmark designed to measure abstract reasoning rather than rote coding. The 25 environments and 183 levels require the agent to infer rules from minimal examples, then apply those rules to novel inputs. If the claim holds, it marks the first time a general-purpose coding agent has achieved a perfect score on this particular challenge.

The improvement from a 30% baseline to 100% is attributed to AVO’s architecture, which wraps a pre-trained model with additional reasoning and execution layers. These layers allow the agent to break problems into sub-tasks, verify intermediate results, and retry failed steps. The cost of adoption is therefore not just the model itself but the orchestration stack that sits around it. Teams would need to replicate or license that stack to achieve similar results.

ARC-AGI-3 is a public set, but the evaluation was conducted internally by Nvidia. No third-party audit or open leaderboard entry confirms the score. Until independent teams reproduce the result, the claim remains provisional. Engineers should treat it as a research milestone rather than a drop-in solution for production workflows.

Even if validated, the agent’s performance on ARC-AGI-3 does not guarantee success in real-world codebases. The benchmark’s environments are self-contained and deterministic; production systems introduce noise, legacy constraints, and ambiguous specifications. AVO’s current architecture may hit limits when faced with incomplete documentation or conflicting requirements.