The diagnosis is instant. The disposition isn't.

A few years back I watched a customer escalation run from first alarm to closed 8D in nineteen days. Detection took eleven minutes. A poka-yoke tripped, the andon went red, the line stopped itself. The other eighteen days were us: the quarantine fence, sorting three shifts of stock, arguing about the escape point, two rounds of verification, signatures climbing the management ladder one name at a time. So when three agent platforms for factory operations launched inside a single week – Snowflake shipping agents with governed business context, SAP positioned by the analysts as a market shaper, NVIDIA open-sourcing a blueprint whose agents coordinate to investigate production issues – I read every press release twice, looking for one thing. Not one of them says what happens to the finding after it is made.

What the blueprint actually buys you

Credit where due: the value is real. Strip the branding and all three offerings converge on the same capability set – autonomous data pulls across MES, maintenance logs and quality records; correlation no analyst with a spreadsheet can match; a first-draft causal narrative in minutes instead of a week. "Torque drift on station 14 correlates with the night-shift tooling change, three days before the escape." I have commissioned exactly that work and paid for it in analyst-weeks, €1,500–€2,500 a week, the narrative landing on Friday afternoon instead of Tuesday morning. Multiply that across every open deviation in a multi-site operation and the arithmetic writes itself.

But notice where every demo stops. At the diagnosis. Diagnosis was never the expensive part.

The disposition bottleneck

QRQC and 8D are not detection systems. They are disposition machinery – containment inside 24 hours, escape-point analysis, the decision to stop the line and notify the customer, verification that the fix held. Every step is owned by a person with a name and something to lose. That is exactly why it works. At SNOP, building the quality function of a 900-plus-person greenfield plant, we took defect costs down 70%, and none of it came from faster detection. Detection was already solved. It came from disposition discipline: who could quarantine stock, who got called at 22:00, how fast containment was verified, who signed D5 through D8.

That machinery runs at human tempo because it is human judgement under accountability. Feed it a queue that delivers finished causal narratives every four minutes and nothing downstream speeds up. The engineers do not get faster. The signature ladder does not get shorter. You have built a high-pressure line feeding a hand-assembly cell.

Here is the question no launch material answers: who signs when an agent names the root cause? An engineer who signs an 8D owns the consequences – contractual, commercial, sometimes legal. An agent can be ninety per cent right and the remaining ten lands on a named human. That asymmetry, not model accuracy, will throttle adoption. From my other life as an ethical hacker: an autonomous agent with write-access to the MES is a persistent identity holding production privileges, with no badge to revoke. The same week, an AI system found four severe vulnerabilities in industrial control software in under four hours. Govern accordingly. NIST has published identity and access tenets for exactly this class of system; read them before the procurement meeting.

A finding without a disposition is just a rumour with a timestamp.

When the agents disagree

I build multi-agent systems in my own time. MultiPS, my orchestration platform, runs 63+ models in parallel with engineered consensus synthesis, and its first lesson was blunt: dissent is not an edge case, it is the resting state. Models disagree about framing, emphasis and conclusions constantly. The engineering work is not eliminating disagreement. It is scoring it, logging it, synthesising around it.

A factory agent stack will be no different. One model reads the data and says tooling wear; another says incoming material lot; a third says both. Most dissent is noise. Some is load-bearing. Under QRQC time pressure the temptation is to grab the first confident narrative and run containment on it – and if containment was built on the wrong cause, you have sorted the wrong parts while the real escape continues upstream.

So treat agent disagreement like a retention sample. Bag it, label it, date it, keep it. When verification at D6 comes back ambiguous, the logged dissent tells you the causal narrative was contested from minute one. An audit trail of disagreement is worth more than a unanimous answer delivered in silence.

The buying test

Before any vendor gets near your MES, run one experiment. Take ten closed 8Ds – real ones, messy ones, including the cases where the first root cause was wrong. Give the agents exactly the data the team had at hour zero, in shadow mode, read-only. Then check three things: whether they reach the recorded root cause, how fast, and where they dissented from each other and from the eventual finding. If the stack cannot reproduce solved history, it has no business near an open escalation. If it can, you have learned something better than any demo – the tempo at which your disposition machinery now has to run.

Two decades of closing 8Ds leave me with one uncomfortable expectation. The factory is about to become very good at knowing what went wrong, very quickly – and no better at all at deciding what to do about it. The diagnosis is instant now. The disposition is not. The plants that win will not be the ones with the sharpest agents; they will be the ones that rebuilt their disposition machinery to deserve them.