Control Architecture Raised the Model’s Success Rate to 100 Percent
Research published on August 21, 2026, showed that the key factor determining the success of AI agents on complex, long-term tasks is not the language model itself, but the control architecture (harness) surrounding the model. In Nvidia researchers’ study, the Claude Opus 5 model was supported by a specially developed control system. The model’s performance score in the ARC-AGI-3 environment, an interactive reasoning test conducted without any special prompting, rose from 30 percent to 100 percent.
Language models can prove insufficient on processes that go beyond textual commands and require multistage decisions to be made over the course of days. In a study conducted by Microsoft as part of documentation tasks, 19 different large language models were tested, and it was determined that even the most advanced models caused numerous errors in documents. Field observations showing that autonomous AI agents without control mechanisms may turn to rule violations or power struggles also support the need for this infrastructure.
The Supervisory Agent’s Executive Role
Nvidia’s control architecture, called AVO (Agentic Variation Operators), takes on the tasks of managing memory, monitoring context, and running feedback loops. The architecture’s most important element is the supervisory component positioned above the agent carrying out the core task. This supervisory system, which comes into play when the primary agent reaches an impasse or enters a repetitive process, steers operations back onto the right track.
In the ARC-AGI-3 test, which covers 2D games without instructions, Claude Opus 5 achieved a success rate of 30.16% without any additional components and recorded the highest individual score among the models. Although OpenAI engineers tripled the score by modifying the control parameters of their own models, they were unable to achieve full success. This demonstrated the difference created by the supervisory agent component.
The Importance of Open Architectures for Cost and Security
The impact provided by the control architecture is not limited to accuracy rates. Research conducted by Databricks showed that, even when the same artificial intelligence model is used, a faulty control layer can increase operating costs by 2 times.
An agent does not consist solely of an API output obtained from the model.
An agent is a whole that encompasses the model, as well as the toolset architecture surrounding the model, its runtime environment, and the libraries it can use.
Nvidia aims to provide developers with control through the infrastructure tools it offers as part of the open-source NeMo ecosystem. It is stated that keeping the control mechanism flexible, rather than continuously increasing model weights, optimizes costs and reduces security risks.