NVIDIA explains how factory AI agents investigate defects and coordinate rework
The early-access FOX design links inspection images, production records and machine readings. Plants must supply equipment connections and decide when operators approve actions.
NVIDIA’s early-access Factory Operations Blueprint (FOX) shows how a manager agent can coordinate specialist agents to assess a flagged circuit board, route confirmed defects for human-led rework, and prepare a root-cause report with recommended corrective action. It is a reference architecture, not a ready-to-run factory product: manufacturers must connect their systems and equipment, set decision and access policies, and validate outputs. Robot transport is shown in simulation, and keeping the workflow on-premises requires every model and service endpoint to stay inside the factory network.
01
NVIDIA published the walkthrough on November 4, following its October 21 announcement of FOX.
02
The visual inspection agent uses NVIDIA Cosmos with factory-approved examples and criteria, without a separate fine-tuning step.
03
For suspected solder defects, the analysis agent can compare oven temperature readings with other production runs and machine manuals.
NVIDIA has detailed how a team of AI agents could follow a suspected circuit-board defect from an inspection alert to a report recommending corrective action. Its November 4 technical post explains the Factory Operations Blueprint, or FOX: an early-access reference architecture that connects factory systems while leaving plants to define which decisions need human approval.
The disclosure follows NVIDIA’s October 21 announcement of FOX. Rather than replace existing factory software, the design adds a manager agent that coordinates specialist agents and equipment. The new walkthrough shows how those pieces work together on an electronics assembly line.
Following one flagged circuit board
The sequence starts when automated optical inspection—a camera-based check—flags a board. That alert is not necessarily a confirmed defect. The manager retrieves the board identifier, affected component and inspection image, then requests another assessment. The identifier lets it connect the image with production records from other stations.
A Visual Inspection Agent calls NVIDIA Cosmos to compare the suspected component with reference examples and approved criteria from the factory handbook. It uses those instructions during inference, without a separate fine-tuning step. An acceptable variation can return to production under factory policy; a confirmed defect proceeds to handling and investigation. An inconclusive assessment goes to human review.
For a confirmed defect, the manager asks a Fleet Management Agent to move the board to rework. That specialist checks permissions and capabilities, then verifies completion. Operators repair and retest the board before packaging. The workflow does not describe agents performing the repair themselves.
The Root Cause Analysis Agent gathers inspection findings, placement records, electrical test results and machine measurements. For a suspected solder defect, it uses the board’s oven entry and exit times to select temperature readings. Kumo Anomaly checks for unusual patterns, which the agent compares with other production runs and machine manuals to investigate whether oven conditions contributed.
Its output is a PDF containing source records, suspect measurements, the likely cause and recommended corrective action for a line manager to review. A failed functional test can also trigger this investigation, even when optical inspection finds no visible defect.
The factory writes the rules
Behind that sequence, the manager uses a language model such as NVIDIA Nemotron to interpret tasks, plan work and assess results. NemoClaw is the default harness: the software connecting it to tools, memory and specialists. OpenShell isolates execution and enforces access policies for files, commands and network destinations.
Factory policies determine permitted actions and when operator review is required. Developers also remain responsible for adapting the blueprint to their production process, including:
Mapping identifiers, timestamps, units and source permissions so records from different systems connect correctly.
FOX supports deployment on an NVIDIA DGX Station with models, data stores and the agent environment hosted locally. But fully on-premises operation requires every model and service endpoint to remain inside the factory network. Choosing an externally hosted model means developers must assess what data reaches it; local hardware alone does not keep the entire workflow local.
Developers can request early access through NVIDIA. The next step is implementation, not simply switching on a finished factory product: FOX supplies workflow examples and integration patterns, while each team supplies its production criteria, equipment connections and output checks.
Sources
developer.nvidia.comHow NVIDIA Factory Operations Blueprint Coordinates AI Agents to Investigate Production Issues | NVIDIA Technical Blog
marketscale.comNVIDIA's FOX blueprint sends AI agents from a defect alert to a root-cause report
Reader comments
Newest comments first. Replies stay oldest first.