The agents agree. The gauge is shared.

A line stops at 06:40 on a Tuesday. By 06:44 the investigation is done: agents have pulled the traceability records, walked the last fifty cycles of process data, tested four hypotheses against history and produced a root cause with a clean narrative arc and a formatted 8D attached. The prose alone would pass any customer audit I have sat through. Here is the uncomfortable observation: when several AI agents converge on the same answer, everyone in the room reads the convergence as proof. Quality people are the only ones trained to distrust a perfect consensus, because we are the only ones who have been burned by one.

What the swarm actually does

NVIDIA has published a factory-operations blueprint in which coordinated agents investigate production issues: a planner decomposes the problem, retrieval agents query the historian and the quality database, an analysis agent tests hypotheses, a writer assembles the report. Siemens argues, meanwhile, that trust must precede autonomy in industrial AI. Both are right about their half. Neither defines the test that earns either. An acceptance test for an autonomous investigator is a harder artefact than a demo, and I have not seen anyone selling one yet.

I have run the smaller version of this experiment. I built MultiPS, an orchestration platform that runs 63 models in parallel under a consensus-synthesis layer. First lesson: consensus is a parameter, not a signal. I tune the thresholds. I weight model families. Weight the similar architectures more heavily and agreement goes up. Nothing in the world has changed. The moment an operator can tune consensus toward an outcome, it stops being evidence and becomes a design choice. A multi-agent investigation is a synthesis engine – five correlated opinions in, one well-formatted opinion out.

Three inspectors, one gauge

Any engineer who has run a gauge R&R under IATF 16949 or AS9100 knows the shape of this problem. The study measures repeatability – same operator, same gauge, same part, run it again – and reproducibility – different operators, same gauge. What it cannot see is bias baked into the gauge itself. Hand one miscalibrated micrometer to three inspectors and they will agree beautifully and be wrong together. That is why no gauge and no NDT technician in a disciplined plant works unsupervised until certified. We studied the measurement before we trusted the measurement.

Map it onto agents.

Re-run the same investigation three times and you have a repeatability study. Run it across different agents and you get something that resembles reproducibility – except the agents are not independent operators. They are inspectors sharing one gauge: overlapping training corpora, the same textbooks, the same failure-mode folklore, the same internet. Correlated bias is the failure mode that makes three agreeing agents worth one.

They were never three measurements. They were one measurement, repeated with better formatting.

Three inspectors passing one gauge back and forth will agree every time. Their agreement measures the gauge, not the part.

The worst root causes I ever chased across automotive and aerospace were the plausible ones. A paint-adhesion defect closed in QRQC in forty-eight hours with a beautifully argued incoming-contamination cause – internally consistent, evidence-fitting, signed. Six weeks later the same defect came back to the same customer, and the second escalation cost roughly four times the first. The real cause was a rinse-tank temperature drift that lived in no log we could pull. That is the trap for agents too. No model can query data that was never recorded, and a model handed incomplete records does not say "I cannot know." It writes a story.

Earning the autonomy

We already run a qualification ladder on suppliers and gauges. Nothing about AI changes the geometry.

• Advisory, 100% review. Every agent conclusion is audited by a human, exactly as a new measurement device is evaluated before release.
• Sampled verification. A defined sampling plan on the agent's conclusions, hit rate tracked per failure class the way you track a supplier's PPM. Autonomy is earned on data, not on a demonstration.
• Bounded autonomy. Pre-approved failure classes only – well-known defect families with verified history. Anything novel escalates to a person.

The investigation itself needs an evidence chain: which records were pulled, which model versions, which prompts, stated confidence – and a named human signature on the 8D. When the customer auditor asks who approved the containment, "the swarm" is not an answer. An unsigned report is a draft.

I have done enough work on the offensive side of security – ethical-hacking certifications, a name on a telecom's bug-bounty list – to recognise the pattern: the same multi-agent architecture is being pointed at industrial targets, and plants are adopting AI faster than they are hardening OT. An autonomous investigator with write access to your MES is an attacker with a lanyard. Bounded autonomy is not bureaucracy. It is containment.

"Trust before autonomy" is the right slogan. In a plant, trust is a documented capability study with acceptance criteria agreed in advance, staged authority and a name on the document – the same discipline we demand of a micrometer, an NDT technician and an auditor. The quality profession wrote that acceptance test decades ago. It is called MSA, staged qualification and a signature on the 8D, and it now applies to the machine that writes the investigation. The plants that win will treat the AI investigator exactly as they treat a new gauge: studied, staged and signed for.