The agents split the work. The defects find the seams.

Here is an uncomfortable observation. The South China Morning Post reports that Chinese factories are putting agentic AI straight into operations – software that schedules, dispatches and adjusts the line rather than advising a person who then clicks approve. The same week, Microsoft talked up coordinated operations for physical AI on the factory floor, L&T Technology Services opened an engineering-intelligence lab around an NVIDIA-accelerated factory, and AZ4C.AI pitched a bundle of agents, data and governance. Half the commentariat is asking whether the West should follow China. Wrong question. Ask the narrower one: when work starts moving agent to agent, where do the defects migrate? Two decades in automotive and aerospace taught me the answer. They migrate to the seams.

The escape always lived at the interface

Pull a real stack of 8D reports – not the sanitised ones – and count how many root causes trace back to a handoff nobody owned. The shift change where a running defect was mentioned but never written down. The engineering release that reached production carrying the old torque spec. The packaging change a supplier made without telling the plant, until the customer noticed first. Interfaces have cost more escapes than any single station I have audited. Stations get control plans. Gaps get assumptions.

I lived this building the quality function of a greenfield automotive plant – 900+ people, new walls, new seams. A stamped bracket reached a customer carrying contamination the customer's line had never seen. The QRQC traced it across three internal handoffs before any single human owned it: press to washing, washing to assembly, assembly to dispatch. Each step assumed the previous one had checked. Each was wrong in the same quiet way.

A defect never escapes a station. It escapes the gap between two stations – each of which assumed the other had checked.

We fixed it the boring way: a name against every interface, a go/look check at each transfer, escalation inside the hour. The quarter we sealed those seams, the plant recorded zero critical customer escalations – the first in its short life – and the defect-cost programme that grew out of it eventually cut cost of poor quality by 70%. None of that came from better stations. It came from owned gaps. Agent-to-agent handoffs are the same topology at a faster clock speed, and nobody has written the control plan for them.

One agent is a signal, consensus is a control

While building MultiPS, the multi-model orchestration platform I designed and still run in private beta, I put the same question to several models in parallel and watched the spread. The first serious run came back with equal confidence and materially different answers – different root-cause categories, different risk rankings, one model flagging a pattern the others had smoothed over.

Nothing was malfunctioning. That was the problem. A single model hands you an answer and never mentions that an equally confident peer disagrees.

So disagreement became the alarm, not the bug. That is why MultiPS runs 63+ models with consensus synthesis: where independent models agree, you have signal strength; where they diverge with equal confidence, you have a free detection event. It is redundant inspection translated into software – two independent inspectors catching what one misses – and the QRQC habit applied to machine judgment. Catch the deviation at the point of deviation, not at the customer.

Write the control plan before the agents clock in

If agents are entering your operations, the countermeasures are not exotic. They are control-plan discipline pointed at software interfaces, and the bill is mostly engineering hours.

• Handoff contracts. Every agent-to-agent transfer gets defined inputs, outputs and tolerances – schema, limits, acceptance criteria. Budget roughly a day of engineering per seam; an escaped defect at a European OEM runs four to six figures in € per incident once sorting, containment and expedited freight are counted.
• Traceability at every transfer. Log the input, the output, the model version, the timestamp. Without that log your 8D dies at "which agent changed it?" – the machine version of the unrecorded shift change.
• Disagreement thresholds wired to reaction plans. Consensus below a set level holds the output, quarantines it and escalates to a human. Reaction plans are control plan 101; they just need wiring into the software layer.
• No agent approves its own output. Maker–checker scales. An agent that generates and verifies itself is an unsealed seam by design.

The agent will not take a sick day, will not take a hint and will not hesitate. Hesitation, in quality, is a feature we have quietly relied on for a century.

China's speed is real and the instinct to match it is understandable. But chasing it with undisciplined agents does not buy speed – it industrialises the escapes. Fast is cheap now; inference is a commodity getting cheaper by the quarter. Sealed seams are the rare thing, and they are a quality discipline, not a procurement decision. The line from advises to acts has been crossed whether the control plans exist or not. Own the handoffs, run the consensus, log every transfer – or the defects will find the seams the way they always have, just at machine speed.