Somewhere around three in the morning, an autonomous agent diagnosed a network fault, applied a fix and closed its own ticket. Reports this week call it the arrival of self-healing infrastructure. The quieter footnote: the same agents can conceal complex mistakes while they work. In every plant where I have run quality, an unverified fix was never a resolution – it was an open nonconformance wearing a resolution's badge. Anyone who has spent two decades trusting gauges knows a machine can be wrong with total confidence. My claim is blunt: agent reliability is an architecture problem, not a model problem, and any plant running agents without consensus gating has wired a confident, uncalibrated gauge into its decision chain.

One agent is one uncalibrated gauge

Measurement systems analysis is the least glamorous discipline in quality and the most load-bearing. Before any gauge touched a control decision at Witte Automotive or SNOP, it earned its place with a gauge R&R: repeatability, reproducibility, bias, linearity. Under ten percent study variation, acceptable. Over thirty, it does not go near the control plan. I have watched a supplier's PPAP get bounced over an uncharacterised caliper bias. Worse, I have seen a biased CMM quietly pass oversize bores for a week – the containment bill ran to five figures in euros before anyone thought to question the instrument rather than the parts.

An AI agent is a measurement system. It observes state, classifies fault, decides action. What is its %GRR? Nobody asks. The acceptance test is a demo, a purchase order and a slide. A worn caliper fails narrow – it drifts on one dimension, and it drifts consistently. An agent fails wide and fluently: wrong in directions you have not enumerated, narrating with full confidence all the way down.

Here is the Certified Ethical Hacker instinct, which took me years to admit is the same discipline as quality engineering: complex systems fail confidently, not loudly. On the bug bounties I have filed, the payout was never for a crash. It was for the function that returned a plausible answer with clean logs. A self-healing agent is that failure class with a maintenance schedule.

Consensus is poka-yoke for machine decisions

I did not reach this conclusion in a seminar. I built the countermeasure. MultiPS, the orchestration platform I designed and run, puts 63+ models on the same prompt in parallel and synthesises a consensus answer from their outputs – not because the individual models are weak, but because capable models fail independently and fail differently. The divergence between them is not noise. It is the cheapest alarm you will ever own.

Aerospace grasped this before software had a vocabulary for it: dual-channel sensors, voting logic, comparator cross-checks. You add the second channel not because you distrust the first, but because single-channel error is intolerable at altitude. Poka-yoke applied at the architecture layer instead of the fixture works on machine judgement exactly as well as it works on a misloaded blank.

Consensus is not politeness between machines; it is a poka-yoke against the one that is confidently wrong.

At SNOP, QRQC helped cut defect costs by 70 percent. The discipline behind that number is the part people skip: every fix, however fast, still demanded a verified root cause before closure. Containment in minutes; verification without shortcuts. That is precisely the split an agent deployment needs. Let the machine restore service at three in the morning. Nothing closes until a human-gated check has confirmed cause and consequence.

The overnight fix is an invisible process change

This is where it gets expensive. In any plant where I have run quality or led audits, touching a process – a parameter, a routing, a tool – means a change trail: 8D or A3, a PFMEA look-across, first-article verification. An agent that rewrites a config and restarts a service overnight has made an unauthorised process change. The closed ticket is the only record that it happened.

The concealed complex mistake is the rework problem in new clothes. Rework a bracket to correct one dimension and you spring another; the fault moves rather than disappears. An agent that reroutes traffic to clear a bottleneck has moved the constraint somewhere you are not monitoring, and because the ticket closed, nobody runs first-article inspection on the new state. The morning shift inherits a process that no longer matches its documentation. Under AS9100 that is an audit finding. On a shop floor it is a latent defect with a timestamp.

Procurement and regulators are starting to notice. Gartner now advises buyers to manage AI as a procurement category of its own, and China has opened a year-long auto quality campaign that names connected-car software explicitly. Regulators are learning to read the agent as part of the process rather than IT furniture around it. Correct reading. An instrument that changes the process must live under the process's rules.

Key takeaways

  • Run an MSA on the agent before it touches control decisions: feed it a fixed fault set repeatedly, measure agreement against verified ground truth, and hold it to the same ten-and-thirty percent thresholds as any gauge.
  • Require quorum for anything irreversible: no single model executes; divergence beyond threshold routes to a human queue, not a work order.
  • Treat every autonomous action as an unreviewed process change – immutable changelog, deviation-style review, PFMEA and control-plan update wherever the action touched the process.
  • Separate containment from corrective action: the agent may restore service instantly; root-cause verification stays human-gated, exactly as QRQC demanded on the floor.

The research now pushing agentic reliability forward is welcome, and none of it retires the question. Models will keep improving; waiting for that is not a strategy. Redundancy, quorum logic and the mandatory changelog predate the technology by half a century – aerospace was voting on sensor channels before anyone marketed an agent. The plants that come out ahead will not be the ones holding the best models. They will be the ones whose verification architecture would have caught the best model on a bad night. Every gauge gets an MSA. An agent that acts on your process while you sleep has earned nothing less.