Nvidia researchers said an agent scaffold called Agentic Variation Operators, or AVO, raised Claude Opus 5's score on the ARC-AGI-3 benchmark from a 30% baseline to a perfect 100%, completing all 183 levels.
The result appears in a technical blog post from Nvidia and was independently reported by TechCrunch. ARC-AGI-3 tests long-horizon, interactive reasoning rather than one-shot question answering, and Nvidia said the underlying model was unchanged between the two scores. The improvement came from AVO's supervisor agent, which watches the main agent's trajectory for stagnation or repeated unproductive cycles and redirects it toward a different strategy when a search plateaus, Nvidia said.
Nvidia also reported AVO completed the public ARC-AGI-3 set using about 12% fewer environment actions, 6,624 total, than a separate system called VISTA needed for the same set, at 7,542 actions. "The model matters, but the model is not the entire agent," Nvidia's post said. TechCrunch reported that OpenAI has separately found that tuning its own harness roughly tripled its ARC-AGI-3 score, though without reaching 100%.
Nvidia vice president Adel El Hallak told TechCrunch the more significant change was adding a supervising agent alongside the agent doing the actual work, rather than any change to prompting or tool access for the primary agent itself.
For builders, the practical read is that a fixed model budget still has headroom: the same underlying model went from failing most of a benchmark to acing it purely through better scaffolding, memory handling and a second agent watching for dead ends. That is worth testing before assuming a harder problem needs a bigger model.