The NVIDIA Factory Operations Blueprint (FOX) is a reference architecture for adding factory manager agents to existing manufacturing systems. Its purpose is to speed up fault investigation, coordinate corrective actions and reduce downtime. A Siemens report cited in the material estimates unplanned downtime costs an average large plant in the surveyed sectors about $253 million per year, underlining the potential business impact.
What FOX does
FOX enables developers to connect current data sources and tools, configure specialist agents, and set rules for when operator review is required. A FOX-based factory manager agent can:
- monitor operations in real time and identify bottlenecks;
- investigate root causes of errors or defects;
- orchestrate robot fleets and autonomous mobile robots (AMRs) to move boards for rework;
- delegate targeted tasks to LLM-based specialist agents (for visual inspection, time-series anomaly detection, fleet management, etc.).
NVIDIA highlights two workflows where these agents add value: production planning and inspection preparation. At the Siemens Electronics Factory in Erlangen (which makes more than a thousand products a year), FOX-built agents combined with NVIDIA Nemotron were used to run simulations through Siemens Intelligence Center X to balance cost, existing orders, supply levels and line setups without a planner manually querying each system. NVIDIA indicates such end-to-end agentic systems can enable up to 20% cost savings. For new product introductions, Siemens combines product lifecycle management and Process Preparation X with NVIDIA Cosmos and NVIDIA Omniverse to create virtual prototypes and synthetic training data for automated optical inspection (AOI), which can shorten time to market by up to 40%.
Architecture and key components
FOX describes how to integrate a factory manager agent with factory knowledge, specialist capabilities and equipment interfaces. Core components shown in the blueprint include:
- LLM (large language model): interprets tasks, links data across systems, develops plans and evaluates results. Examples listed are GPT-6 Astra (frontier model) and NVIDIA Nemotron (open model option).
- Agent harness: connects the manager to tools, persistent memory and specialist agents; NemoClaw is the default harness.
- NVIDIA OpenShell: isolates execution and enforces access policies for files, commands and network destinations. Factory policies determine allowed actions and when human review is required.
Developer responsibilities include defining tasks and completion criteria, integrating applications, onboarding data and mapping identifiers, timestamps and units, handling equipment integrations, choosing resources for deployment and configuring model endpoints.
Data connections and specialist agents
FOX connects machine databases, manufacturing execution systems (MES) and enterprise resource planning (ERP) systems via adapters and mappings. Manuals and operating procedures help the manager interpret records. An Elasticsearch database stores evidence and preserves investigation context for retrieval.
The manager delegates focused work to LLM-based specialist agents, each with its own instructions, context, skills and tools. Specialist agents can call models such as NVIDIA Cosmos for visual inspection and Kumo for time-series anomaly detection, then return findings that guide the manager’s next actions. Developers adapt these roles to their production processes.
On-premises deployment
FOX can run fully on-premises on an NVIDIA DGX Station with the agent environment, models and data stores hosted locally. In a fully on-premises setup every model and service endpoint remains within the factory network so that production logs, images, prompts and investigation outputs are kept local. If developers use an externally hosted model, they must assess what data is sent to that endpoint.
SMT (surface-mount technology) reference example
The blueprint presents a reference example for an SMT assembly line producing printed circuit boards (PCBs). SMT automates most assembly steps, but faults found at one station may originate earlier in the line. Typical SMT stations listed are solder paste printer, pick-and-place machine, reflow oven, AOI and in-circuit testing (ICT), board integration, SOP station for manual tasks, functional test and packaging.
When an AOI alert occurs, it might be a false positive; a confirmed defect or a failed functional test can require rework. In the reference, FOX coordinates inspection, handling and investigation as follows:
- Receiving the AOI alert: the AOI station exposes inspection records to the manager via an MCP server or API. The manager retrieves the board identifier, the affected component and the inspection image, then requests a secondary assessment.
- Linking processing history: the board identifier connects to records from other stations; adapters preserve identifiers, timestamps, units and source references so the agents can retrieve the board’s production history and machine conditions.
- Visual Inspection Agent: using the image, reference examples and approved criteria, the agent calls NVIDIA Cosmos to compare the suspected component against the factory handbook. Cosmos can use in-context learning during inference without separate fine-tuning. The specialist returns findings, evidence and any uncertainty; acceptable variations are passed back as support for continued production, confirmed defects proceed to handling, and inconclusive results go to human review.
- Fleet Management Agent: for a confirmed defect the manager asks the fleet agent to move the board to rework. The reference demonstrates this transport in a digital twin with NVIDIA Isaac Sim and ROS 2. Physical deployment requires factory-specific integrations for robot handling and AMR transport; initial deployments should require operator approval before handling. Operators then repair and retest the board.
- Root Cause Analysis (RCA) Agent: using the board identifier and processing history it retrieves inspection findings, placement records, electrical test results and machine measurements. For suspected solder defects the agent examines reflow oven entry and exit times and uses Kumo Anomaly to analyze oven readings for unusual patterns. The agent compares findings with other production runs, inspection results and machine manuals. Elasticsearch provides access to production evidence, versioned manuals and prior investigations. The RCA agent returns a PDF report containing source records, suspect measurements, the likely cause and recommended corrective action for the line manager to review.
Benefits and customization
In the SMT example FOX coordinates inspection, root-cause analysis and fleet management through connections to existing databases, factory systems and machines. The architecture reduces manual effort, helps trace defects to probable causes, coordinates rework and supports decisions to prevent further failures. Developers can customize agent skills, tools and specialist roles as factory needs evolve.
Getting started
NVIDIA FOX and the NVIDIA Agent Toolkit help developers build factory manager agents that connect evidence across systems, investigate issues and coordinate corrective actions, aiming to reduce downtime and improve quality while leveraging existing factory software. FOX can be requested for early access, and NVIDIA provides resources for building specialized agents with VSS.
