The agent standard shipped. Your audit trail didn't.

The machines on your line now have a safety standard with a release-notes page. I have spent two decades watching standards get born the slow way – IATF 16949 revisions crawling through committee drafts, AS9100 changes argued clause by clause, VDA 6.3 rating debates that outlived the auditors who started them. Last Friday a foundation-model company shipped an open hardware standard for AI agents that command physical machines, safety guidance included. It shipped between two Fridays. Your certification body has never heard of it. Your systems integrator will be quoting it by Q4, and someone in your steering meeting will accept it as mitigation.

What the standard covers – and what it quietly doesn't

Anthropic's Model Hardware Standard is a genuine first: a de facto control layer for machine autonomy, published before ISO, IATF or SAE have even convened a working group. Credit where due. Somebody sat down and thought about agents closing valves and driving actuators before an incident forced the question. Now read it the way you read any supplier document – for what is absent. No traceability obligations. No change-control hooks into your management-of-change process, no audit-trail requirements, no record retention, no duty to log which model version issued which command to which machine at which time. An agent that clears a fault overnight and quietly buries a complex mistake is, under this document, a fully conforming agent.

It reads like a demo document because vendor documentation always describes the demo, never your line. I design and deploy multi-model agent systems; the platform I built runs 63 models in parallel with consensus synthesis, precisely because one model's confident answer is not evidence. I do not train models. I integrate them, and integration teaches you fast that the distance between the vendor's test rig and your press line is where quality lives or dies. Open-source safety guidance means everyone can read it. Three people will.

A standard you cannot audit is a nonconformance generator with a press release.

When a vendor standard meets IATF 16949 and AS9100

Two architectures now share one line. Your certified QMS demands documented control of every process affecting product – process controls, reaction plans, records an auditor can follow from a customer complaint back to a machine parameter. The agent executes machine commands under a specification your certificate does not recognise and your control plan does not mention. Neither document yields. When the vendor standard says an autonomous interlock is acceptable and your control plan says an operator gate is mandatory, the deviation is yours. Your name is on the certificate.

I once walked an auditor from a customer complaint back to a single torque parameter on a Saturday shift, record by record, and the finding closed because the trail held. Run that same walk with an agent in the loop and the trail stops at the API boundary. That is not a technicality; that is where the nonconformance gets written.

I know what a standard costs to implement, because I have paid it. At SNOP I built a quality system for 900+ people from zero at a greenfield plant – every clause earned an owner, a record, a training plan and an audit trail before it earned trust. Nothing real arrives between two Fridays. So define ownership now, in writing: which role arbitrates when vendor guidance and the control plan disagree on the same line, who signs the deviation, who faces the customer. Because when the 8D lands on your desk, "the agent was compliant with the vendor standard" is a root-cause narrative, not a corrective action. No auditor on earth accepts conformance to a document they cannot sample.

The supplier-documentation playbook

Treat it exactly like a supplier-provided specification, whatever the press coverage implies:

• Verify it against your own failure modes. Map every agent-authorised action to your PFMEA. Where the vendor assumes a failure cannot happen, your process history disagrees.
• Validate it on your line, not the vendor's rig. Two weeks on your own equipment costs a few thousand euro. One escaped critical defect burns that per hour, plus the customer.
• Keep a human gate at every state change on safety-relevant and product-characteristic equipment. Autonomy earns scope; a datasheet does not grant it.
• Pin and log everything. Model version, prompt, command, machine response, timestamp – immutable, agent-uneditable, attached to the device history record.

Hard-won rule from deploying multi-model systems: single-agent output is a claim; consensus output is a control. When a decision changes the state of a machine, you want agreement between independent models – or an independent verification step – behind it. Not one fluent answer.

Standards used to flow upward – from plants, from auditors, from failure investigations, argued into committees over years. Now they flow downhill, from model vendors, between release notes. The direction changed. The accountability did not. Quality can take a seat at that table while the document is still soft enough to argue with, or spend the next decade writing 8Ds against rules nobody in the plant ever read. I have spent twenty years cleaning up after the expensive kind of shortcut. I am not volunteering to fund another decade of them.