AVO runs an iterative inspect-plan-implement-evaluate loop backed by two mechanisms for sustaining long-horizon work: persistent memory that carries forward prior implementations, evaluation results, and tool output across context-window boundaries, and a supervisor that monitors the broader search trajectory for stagnation and can redirect the main agent when progress stalls. NVIDIA first built AVO for autonomous GPU-kernel optimization on DGX B200 systems, then applied the same unchanged core loop to ARC-AGI-3, an interactive reasoning benchmark, reporting a 100.00 RHAE score across the public set with only the task interface swapped.