On August 27, Anthropic opened a research preview of the Model Hardware Standard (MHS) — a specification that lets AI agents discover, understand and safely operate physical devices. The launch partners are mostly research labs, but the vendor list (Universal Robots, Doosan, Tecan, QIAGEN, Raspberry Pi, LeRobot) is what made me sit up.
I’ve sat through enough IT/OT architecture meetings to know how „yet another integration standard“ usually lands. So I wrote a longer piece that does three things:
- Explains what MHS actually is — and what it isn’t. Spoiler: it’s not a fieldbus, not a control protocol, and not a PLC replacement. It lives above all of that, roughly between ISA-95 Levels 2 and 3.
- Puts it next to OPC UA, MQTT and MTP. The three (four, if you count SiLA 2) solve different problems, and the overlaps are exactly where the friction will be. My view: MHS should be generated from OPC UA information models and MTP files, not written beside them. Two self-descriptions of the same machine will drift apart.
- Asks what it means when an AI company writes the standard. The precedent is MCP, which went from release to de facto standard in about a year. Manufacturers waiting for an IEC version may find the market has already decided. At the same time: safety, cybersecurity, GxP validation and single-company governance are real concerns, not footnotes.
The one pattern I’d want the industry to insist on: the agent explores and experiments, then hands over a deterministic, inspectable script that runs without an LLM in the loop. That’s the right architecture — and the case studies show it works.
Read the full article on LinkedIn: An AI Company Just Published a Hardware Standard. Manufacturers Should Pay Attention.
I’d be curious to hear from people on the OPC Foundation, VDMA or NAMUR side — the spec is still malleable, and now is the time to be in the room.
