The device picks the regime - Anthropic's hardware standard, and the deadline that is not 2028

Anthropic previewed the Model Hardware Standard on 27 August 2026 - a driver specification that lets an agent operate microscopes, liquid handlers, robotic arms and lasers through read and write primitives, standard device discovery and safety limits declared in the driver and enforced at the device, with QuEra's optical lock recovery moving from 58 per cent at about 150 seconds to 99.3 per cent across 700 trials. The obvious reading, that this is MCP for robots, misses the classification: Article 6(1) makes an agent enforcing a device's safety limits a safety component of an Annex I product, and the Digital Omnibus then moved Regulation (EU) 2023/1230 from Section A to Section B of Annex I, so that under Article 2(2) only Article 6(1), Article 60a and Articles 102 to 112 apply and the high-risk requirements arrive inside the Machinery Regulation by delegated act by 2 August 2028. The consequence is a date correction: the Machinery Regulation itself applies from 20 January 2027, eighteen months earlier, and its Annex I reaches safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions, which is a notified body rather than a self-declaration - so the practitioner moves are to classify the device before wiring the agent, log who set each safety limit and what enforced it, ask the hardware vendor in writing for its January 2027 conformity route, and keep the driver's natural-language capability file as the intended-purpose document it already is.

Post

Anthropic wired Claude to a laser last week. The compliance question that opens is not the one everyone is asking.

On 27 August, Anthropic previewed the Model Hardware Standard — a driver spec that lets an agent operate physical instruments: microscopes, liquid handlers, robotic arms, lasers. Read and write primitives, device discovery, safety limits declared in the driver and enforced at the device.

The numbers are Anthropic's own. At QuEra Computing, recovering an optical lock started at 58% and about 150 seconds. It ended at 99.3% across 700 trials — 695 recoveries. Over a 19-hour run, Claude's PID controllers never lost the lock; the expert-tuned ones dropped it about 1.6 times an hour.

Three things before you file this under "MCP for robots".

  1. It makes the agent a safety component. AI Act Article 6(1): high-risk where the system "is intended to be used as a safety component of a product" covered by Annex I, and that product needs third-party conformity assessment. An agent enforcing a device's safety limits is that, by description.

  2. Then the Act steps back. The Digital Omnibus moved Regulation (EU) 2023/1230 — machinery — from Section A to Section B of Annex I. For Section B, Article 2(2) leaves only Article 6(1), Article 60a and Articles 102 to 112 applying. The high-risk requirements arrive inside the Machinery Regulation instead, by delegated act, by 2 August 2028.

  3. Which means your calendar has the wrong date. AI Act Annex I high-risk: 2 August 2028. The Machinery Regulation itself: 20 January 2027. Eighteen months earlier — and its Annex I reaches "safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions", which is a notified body, not a self-declaration.

Monday:

  1. Classify the device before you wire an agent to it. A robotic arm is machinery, Section B. An IVD analyser is Regulation 2017/746, Section A, and the whole high-risk chapter comes with it.
  2. Log who set each safety limit and what actually enforced it — driver, firmware or model.
  3. Ask your hardware vendor in writing which conformity route it plans for January 2027.
  4. Keep the natural-language capability file the driver generates. It is your intended-purpose document whether you meant it or not.

The standard made the wiring cheap. It did not move the responsibility.

Which of your devices decides your regime?

#EUAIAct #AgenticAI #MachineryRegulation #AICompliance #Robotics

First comment

Sources, primary first.

Anthropic's own announcement of the Model Hardware Standard, 27 August 2026 — the QuEra, Genentech, Carnegie Mellon and Janelia numbers all come from here: https://www.anthropic.com/news/model-hardware-standard-research-preview

EU AI Act Article 6(1), the two conditions that make an AI system high-risk by product route: https://artificialintelligenceact.eu/article/6/

Article 2(2) — what is left applying when the product sits in Section B of Annex I: https://artificialintelligenceact.eu/article/2/

Annex I, Sections A and B. Machinery is item 21, Section B; in vitro diagnostics is item 12, Section A: https://artificialintelligenceact.eu/annex/1/

Eurogip on what the Digital Omnibus changes for the Machinery Regulation, including the Section A to Section B move and the 2 August 2028 delegated act: https://eurogip.fr/en/machinery-regulation-2023-1230-what-the-digital-omnibus-on-ai-changes/

Regulation (EU) 2023/1230 on machinery — applies 20 January 2027: https://eur-lex.europa.eu/eli/reg/2023/1230/oj/eng

Cut for length: MHS does not yet work with hardware that has no programming interface, and Anthropic is explicit that spatial and physical reasoning "still require expert oversight" — Claude struggled with bubble formation in liquid handling. Which is the honest read on Article 14 human oversight: the vendor is telling you where the human still has to stand.

Claims

Sources

Corrections

No corrections. Three checks fired during research and are recorded here because each changed the edition:

  • The aggregator sweep dated the Model Hardware Standard to 31 August 2026; Anthropic's own post is dated 27 August 2026. The event date is used.
  • The same sweep listed a NIST agentic-identity item as 31 August news; the underlying NIST concept paper is from March 2026. Dropped.
  • QuEra's figures span stages: 96 per cent at about six seconds is an intermediate result, 99.3 per cent across 700 trials the endpoint. The post compares the start (58 per cent, about 150 seconds) with the endpoint and says so.

AI disclosure

The video for this edition contains AI-generated media: the narration is a synthetic voice and the cover art is generated. The script, the claims and the source verification are human-authored and human-checked. Assets carry IPTC DigitalSourceType and C2PA Content Credentials.