NVIDIA's FOX blueprint automates factory defect workflows

From alert to report

NVIDIA recently introduced the Factory Operations Blueprint, or FOX, which serves as a reference architecture for factory agent workflows. The system is designed to handle the complex coordination required when a production line identifies a defect. Instead of a human having to manually gather data from multiple sources, FOX uses a hierarchy of AI agents to manage the process.

In a typical example, an automated camera flags a circuit board defect on a surface-mount line. A Visual Inspection Agent performs a second check, while a robot moves the board to a rework station. Meanwhile, a Root Cause Analysis Agent scans machine data to find the source of the error and generates a PDF report with a recommended fix.

The goal is to bridge the gap between different systems like machine databases, enterprise resource planning (ERP), and manufacturing execution systems (MES). As MarketScale reported, this approach helps teams turn available data into a response much faster than manual investigation allows.

How the agent hierarchy works

The FOX architecture relies on a central manager agent that sits on top of the existing factory systems. This manager agent runs on a large language model, such as NVIDIA Nemotron, and uses a harness called NemoClaw to interpret tasks and maintain persistent memory. It acts as the brain that directs various specialist agents to perform specific jobs.

Each specialist agent has a distinct role tailored to a part of the production chain. For instance, the Visual Inspection Agent uses a world understanding model to compare components against a handbook without needing to be retrained. The Root Cause Analysis Agent scans time-series data, such as oven temperatures, to find unusual patterns that might cause defects.

Additionally, a Fleet Management Agent coordinates the physical movement of parts using robotics frameworks. By isolating these tasks into specialized agents, the system can handle high-mix production environments where updating a single model for every new product would be too difficult. Instead, the agents can read new defect criteria from documentation as a simple data task.

Configurable human oversight

A key feature of the FOX blueprint is how it handles human intervention and operator review. Rather than creating different versions of the software for different needs, NVIDIA makes approval rules a configurable setting. Plant managers can write into the configuration exactly which agent decisions require a human sign-off before an action is taken.

This flexibility allows different types of facilities to use the same blueprint. A quality engineer might want every single rework dispatch approved by a person, while another plant might allow the agent to move boards autonomously but hold back any changes to the production process. This design treats human oversight as a preference rather than a technical limitation of the product.

NVIDIA suggests that starting with read-only investigations is a good way to build trust with floor workers. By showing the reasoning and evidence behind an agent's conclusion, engineers can verify the data before the system is granted the authority to execute physical actions. This approach helps move the technology from a laboratory setting to a live production environment.

Real-world results and deployment

NVIDIA has already begun integrating these technologies into industrial settings, specifically with Siemens. At the Siemens Erlangen site, order-planning agents built on the FOX framework have enabled cost savings of up to 20%. These agents run simulations that weigh costs, supply levels, and line setups so planners do not have to query each system by hand when a rush order arrives.

The blueprint also supports fully on-premises deployment, which is critical for many industrial manufacturers. This means that every model, image, and prompt stays inside the factory network rather than being sent to the cloud. This setup ensures that sensitive production records and proprietary data remain secure while still benefiting from high-performance AI.

While FOX is currently in early access as a developer blueprint, it is built on components that have already handled production loads in NVIDIA's own data centers. The same underlying stack manages GPU clusters, correlating causes across hardware and network layers. This experience provides a foundation for the factory workflows now being offered to systems integrators and factory software teams.