NVIDIA's FOX blueprint puts a factory manager agent in charge of inspection, rework and root-cause agents

Factory operations gain a coordinator: a factory manager agent works on real-time factory data to direct specialist agents and existing equipment, tying cross-system evidence gathering and response into one flow. Previously, whether investigating a failed circuit board or responding to a rush order, factory teams had to gather evidence across systems, consult specialists and coordinate the next step; the data was available, but turning it into a response took substantial time and effort. FOX's SMT reference example follows one AOI alert through inspection, rework and root-cause analysis, showing that the same architecture can take different data, specialist agents and equipment connections.

The manager agent is built from three parts: a large language model that interprets tasks, connects data across factory systems, develops plans and assesses results; an agent harness that connects it to tools, skills, persistent memory and specialist agents; and OpenShell, which isolates the execution environment and enforces access policies for files, commands and network destinations. The harness defaults to NemoClaw with support for alternatives, managing background tasks and scheduled jobs while keeping the manager available to the operator; the model can be a frontier model such as GPT-6 Astra or an open model such as NVIDIA Nemotron. Developers connect existing machine databases, manufacturing execution systems (MES) and enterprise resource planning (ERP) systems through adapters and mappings; manuals and operating procedures help the manager interpret the data, and an Elasticsearch database stores information for retrieval and retains investigation context.

Specialist agents take focused tasks with their own instructions, context, skills and tools, then return findings that guide the manager's next step. The Visual Inspection Agent calls the world understanding model NVIDIA Cosmos to compare a suspected component against the factory handbook, using in-context learning to apply examples and instructions at inference without a separate fine-tuning step; the Root Cause Analysis Agent uses Kumo Anomaly to analyze unusual patterns in reflow oven readings. Inspection outcomes split three ways: an acceptable variation returns evidence supporting continued production under factory policy, a confirmed defect proceeds to handling and investigation, and an inconclusive assessment goes to human review; for a confirmed defect the manager asks the Fleet Management Agent to move the board to rework, demonstrated in a digital twin using Isaac Sim and ROS 2.

The architecture targets developers who want to add a coordination layer to existing factory software; the SMT reference example demonstrates rework transport in a digital twin, and physical deployment requires factory-specific robot handling and autonomous mobile robot transport integrations.

The material states that the SMT reference example is separate from the Siemens production-planning implementation; initial deployments should require operator approval before handling; and developers using an externally hosted model must assess what data is sent to that endpoint.