Nvidia’s Agentic Variation Operators (AVO) is a research harness that wraps an AI model with tools, memory, feedback loops, and a supervisory agent. By placing Claude Opus 5 inside this system, Nvidia lifted its score from 30% to 100% on ARC-AGI-3, solving all 183 puzzles across the benchmark’s 25 interactive games. The result suggests that harness design can matter as much as, or more than, changing the underlying model for difficult, long-running agent tasks.
Source