Nvidia has published research showing that the software surrounding an AI model can dramatically change how well it performs on long-running autonomous tasks. Using its Agentic Variation Operators, or AVO, system with Claude Opus 5, Nvidia achieved a 100% score on the public ARC-AGI-3 interactive reasoning benchmark, compared with a roughly 30% reported baseline for the model.
Nvidia detailed the results in an official research post. The company argues that an AI agent depends not only on the underlying model, but also on its harness, which manages tools, memory, context, feedback, and how the system recovers when it gets stuck.
AVO Uses Persistent Memory and a Supervising Agent
ARC-AGI-3 consists of unfamiliar 2D game environments with no instructions or stated objectives. An agent must experiment, determine how each environment works, remember what it learns, and complete increasingly difficult levels.
Nvidia’s AVO system completed all 183 levels across 25 public environments with a 100.00 Relative Human Action Efficiency score. It used 6,624 environment actions, about 12% fewer than the 7,542 actions reported by VISTA using Claude Opus 5.
A key part of AVO is persistent memory, which carries information from previous actions and results forward rather than forcing the model to repeatedly reconstruct its understanding. It also uses a supervising agent that monitors progress and can redirect the main agent when it repeats unproductive approaches or reaches a dead end.
Nvidia AI vice president Adel El Hallak compared the supervisor to a CEO that can guide an agent back toward a productive direction.
Other Research Has Found Similar Harness Effects
OpenAI recently reached a related finding while testing its models on ARC-AGI-3. By changing two harness settings, the company tripled its benchmark performance, although its models still remained below Nvidia’s reported AVO result.
Long-horizon work remains difficult for current AI systems. Microsoft research published in April tested 19 large language models on extended document-editing tasks and found that all of them introduced errors.
Harness design can also affect cost. Databricks research found that the same AI model can produce significantly different costs depending on the agent system surrounding it.
Nvidia developed AVO as a research architecture rather than a new commercial product. The company also provides agent-building components through its NeMo software ecosystem, while the AVO research describes how the system uses memory, supervision, tools, and repeated evaluation to sustain work over longer periods.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.