The Mud King, home

Agentic AI · Manufacturing

Agentic AI on the factory floorAgents, knee-deep in the plant

Factories already have the data: sensors, SCADA, MES, ERP. What they lack is someone who acts on it in time. I built the Digital Factory Console to show what changes when an agent makes that call, inside the controls a plant needs.Your plant already knows the lot is bad. The camera said so six times. This is what happens when something finally listens.

Demonstration build

Open the live console (opens in a new tab)

A simulated plant with illustrative numbers, not a client system. Every visitor shares the same plant, and it resets daily. The console is in English.

Same defect, two lines

The lot that ships, and the lot that stays.Six alarms. One shipment. Four hours of regret.

The console’s first screen runs one plating defect on lot L-1045, 2,400 pin headers on a China line, twice: once as most lines run today, once with agents. The figures are the simulation’s.

Today

minutes to contain the defect
250
minutes plate press P2 is down
38
of this lot left the plant
$432

The press current is already in the historian. The camera raises six alarms. The operator clears them, because the lot is late, and L-1045 leaves the line. Containment starts after the shipment.

With agents

minutes to contain the defect
4
minutes plate press P2 is down
0
of this lot left the plant
$0

The same drift reaches a repair check: an anode basket, and a plater on this shift. Stock is zero, so no work order opens and the press keeps running. The six alarms become one incident at Inspect. The agent drafts a hold, a second person approves it, and the lot stays.

Where the agent sits

The apps report facts. The agent decides.Five dashboards, zero decisions. Until now.

Computer vision, predictive maintenance, inventory, planning and OEE each report one fact. The agent is the role that uses those facts on one lot, in one decision.

  • Computer vision

    The camera sees a plating void on pin 3.

    Turns six alarms into one incident at Inspect.

  • Predictive maintenance

    Press P2’s current is drifting. The drift is the cause.

    Asks for a repair only when a plater is on this shift.

  • Inventory

    The anode basket the repair needs is out of stock.

    Leaves the work order closed, so the press keeps running.

  • Planning

    The lot is late, and the schedule wants it out the door.

    Holds it at Inspect, so the defect is settled first.

  • OEE

    Press P2 is the line’s constraint.

    Keeps its downtime at zero while the repair cannot be done.

  • Closed-loop quality

    A hold stops the lot, and a release lets it go.

    Drafts the hold. A second person approves it, and release uses the same rule.

The digital thread underneath

An agent is only as good as the plumbing.No pipes, no agent.

Agents cannot act on data that never leaves the machine. Behind the line screen, the console runs the full stack, 13 systems and 21 live data flows, and lets you switch any system off to see what breaks.

  1. Shop floor and edge

    Sensors and PLCs, edge gateway, SCADA, historian

  2. Execution

    MES, manufacturing operations (MOM), SAP ERP

  3. Data platform

    Lakehouse, semantic layer, an ontology of machines, orders, batches and materials

  4. Applications

    Digital twin, machine-learning forecasts, dashboards, agents

The controls

Autonomy, inside the rules of the plant.It can act. It can’t go rogue.

The same rules I write into an agent’s job description, built into the console, so you can see them work.

  • Two-person approval

    Safety-relevant actions, such as stopping a machine or holding a lot or an order, wait for a second person. The approval is recorded.

  • Roles

    Operator, engineer, plant manager and executive each see their own view. Every switch between roles is logged.

  • Audit log

    Every system event, user action and agent action is recorded with time, actor, action and outcome.

  • Autopilot, with limits

    Closed-loop recovery restores failed systems, connectors and machines on its own. Safety and AI actions still need a person.

  • OT security

    IEC 62443 zones from field devices to enterprise IT, and an attack lab with six scenarios, including prompt injection against an AI agent, run with and without protections.

Inside the console

What you will seeScreenshots, for the skeptics

Screens from the live console, which is in English.

Console line view in Today mode: 250 minutes to contain the void, plate press P2 down 38 minutes, $432 of lot L-1045 shipped.
Today: the operator clears six alarms and the lot ships.
Console line view in Agents mode: 4 minutes to contain the void, press P2 down 0 minutes, $0 of the lot shipped, with an “Approve hold, second person” button.
With agents: one incident, a held lot, a second person to approve.
Console data-flow view: sensors, edge, SCADA and historian feeding MES, MOM and SAP, then a lakehouse, semantic layer and ontology, then digital twin, ML and dashboards.
The digital thread: switch any system off and watch the impact.
Console CyberAttack Lab: six attack scenarios, eleven protection toggles, and minutes to recovery with and without protections.
The attack lab: the same attack, with and without protections.

Open the live console (opens in a new tab)

Where this comes from

Built from programs I have run.I’ve stood in this mud before.

The console is a demonstration. The patterns in it come from programs on my record.

  • Project Phoenix linked SAP S/4HANA, IIoT, MES and digital twins across 45+ production lines. OEE rose 32%, and the Suzhou factory joined the World Economic Forum’s Global Lighthouse Network.

    Read the case study
  • Digit-AI took AI use cases from pilot to production, with a 17.2% efficiency improvement across an industrial group’s China business. The same program put plant-edge computer vision for quality control on the line at Tianjin.

    Read the case study
  • One security platform for IT, OT/ICS and AI security governance, aligned to IEC 62443 and covering 170+ regulations in 60+ countries.

    Read the case study

Have a line in mind?Got a line that keeps shipping its mistakes?