The new achievement demonstrates that agent ecosystem engineering and intelligent management are the decisive elements in achieving independent reasoning and solving previously unseen problems.
Article index:
- The software architecture that made the technical difference
- From GPU acceleration to general reasoning
- The impact of agent architecture on AI costs
- Frequently asked questions
The software architecture that made the technical difference
On Friday, Nvidia published a research paper revealing that its general-purpose autonomous agent system, known as “Agentic Contrastive Operators” (abbreviated as AVO), successfully achieved an unprecedented 100% score on the advanced interactive reasoning benchmark “ARC-AGI-3”. This remarkable result is definitive proof that agent architecture and runtime environment engineering, rather than foundational model capabilities alone, are the decisive factors in determining performance levels when executing complex autonomous tasks.
Using the Claude Opus 5 model as the foundational reasoning model, the AVO system successfully passed all 183 levels distributed across 25 interactive environments within the benchmark’s general test suite, achieving a score of 100.00 on the relative human action efficiency scale and completing the entire task in just 6,624 environmental actions. In contrast, when running the Claude Opus 5 model alone without Nvidia’s agent software environment, the model achieved only 30% on the same benchmark, despite being the highest score recorded among all models tested at high reasoning effort levels.
Considered one of the most difficult artificial intelligence tests, the ARC-AGI-3 benchmark consists of two-dimensional interactive environments resembling games but entirely devoid of any instructions, rules, or stated objectives. This forces the intelligent agent to explore the environment independently, deduce movement dynamics and rules, and then act efficiently across progressively more difficult and complex levels. This benchmark has posed an elusive challenge for leading models, with OpenAI models scoring below 10% before the company conducted its own research to improve agent environments last month.
喜び (AVO) relies on two main features that make a difference: the first is continuous and permanent memory that transfers software applications and previous reasoning experiences across successive sessions, and the second is the presence of a supervisor component that monitors the agent’s workflow and intervenes to redirect it as soon as it notices stagnation or repetition of unproductive paths. Adel Hallak, vice president of products in Nvidia’s AI unit, said in statements to TechCrunch: “The most interesting aspect was introducing a supervisory and guiding agent alongside the main agent executing the actual work,” comparing the supervisor’s role to “a CEO who provides the necessary directional pushes whenever the agent strays off the correct path.”
From GPU acceleration to general reasoning
The AVO system was originally developed to optimize GPU processing cores, running continuously for seven consecutive days. During this time, it explored more than 500 directions and paths for software optimization, successfully generating processing cores that outperformed the cuDNN library by up to 3.5% and the FlashAttention-4 library by up to 10.5% on advanced Nvidia DGX B200 systems.
This innovative architecture successfully transitioned to the general reasoning benchmark without needing to modify the core CPU processing loop, with changes limited solely to environment-specific tools and evaluation criteria. The system also demonstrated exceptional flexibility across multiple artificial intelligence models, as limited experiments where AVO was integrated with the GPT-5.6 Soul model showed the model reaching required levels in real-time faster in several cases, while the Opus model consumed fewer environmental actions.
The impact of agent architecture on AI costs
Nvidia’s recent research adds strong evidence to a growing body of scientific literature confirming that the agent runtime environment and its engineering, rather than simply choosing the intelligent model, is the decisive variable in autonomous systems efficiency. Databricks CEO Ali Ghodsi noted in July that choosing the agent runtime environment can double AI costs for the same model, pointing out that superficial evaluation of cheap and expensive models overlooks the most important factor: the nature of the agent architecture used.
Hallak concluded by emphasizing that these results support Nvidia’s vision to build open software agent packages, explaining: “We believe that owning an open agent package that gives you full control over the runtime environment, infrastructure, and execution time is the primary requirement to push the AI ecosystem forward safely and efficiently.”
Frequently asked questions
Question: What achievement did Nvidia’s AVO agent system accomplish?
Answer: It achieved a perfect 100% score on the ARC-AGI-3 interactive reasoning benchmark across 183 levels.
Question: What architectural components helped the system achieve this result?
Answer: They include continuous memory across sessions and the presence of a supervisory agent that monitors the path and redirects the executing agent.
Question: How does the agent runtime environment affect the efficiency and cost of intelligent models?
Answer: The runtime environment determines task execution efficiency, and poor choices can double computing costs.
