Anthropic's hardware standard moves agent safety onto the wiring

A commenter on Ars Technica's forum thread describes his daughter pointing Claude at a LeRobot arm late last year, with three cameras as a feedback loop: the agent wrote scripts to move the arm to preset positions, filmed the results, and used the camera and position data to train a model that generalized pick-and-place, all of it done over Signal without her touching the codebase ^[14]. That is the deployment pattern worth holding in your head, not the Janelia microscope rig. Physical control arrives through whoever finds the driver convenient.

The hardware layer used to come with ceremony: a procurement review, a safety officer, an interlock inherited from the machine's original vendor. A standard driver removes the integration work that used to function as a delay, and the delay was doing safety work nobody had formally assigned to it. Anthropic's own pitch is that setup running weeks or months collapses to hours or minutes ^[4]. Take the ends of that range literally and it is about 84x at the conservative end, two weeks against four hours, and roughly 1,440x at the aggressive end, two months against one hour ^[12]. Whichever number is real, the friction being removed is the friction that gave somebody time to ask what happens if this goes wrong.

Anthropic mentions one detail almost in passing: MHS needs no model at all; devices can be driven directly from command-line prompts and API code files ^[7]. The model is an optional consumer of the interface, which means the driver layer is also where control can sit. It has to sit somewhere outside the model, because the sentence Anthropic uses to sell the model integration is that models can update parameters in real time and, in some cases, recover from hardware errors without intervention ^[8]. Stopping one of the laser beams in Bast's rig takes a physical intervention that a filter operating on tokens cannot supply ^[6]. The reported description of MHS covers drivers and setup time, not an emergency stop or a physical interlock ^[11].

So the inventory that matters is not the device list. Two questions give you four boxes. Is the action reversible? Does stopping it require the model's cooperation? Reversible plus a stop that works without software is the only box that runs unattended. Irreversible plus a software-only stop is the box where you owe someone a physical break in the circuit before the device gets an MHS driver at all. The two mixed boxes need a named person rather than a written policy. Commenters in the thread reached for wire cutters and a four-pound mallet ^[10], which is crude and also the right layer.

The pressure runs the other way, though. The same thread carries a commenter's read that AWS treats human-in-the-loop as expensive, because it pauses AI actions while a person reviews and approves them ^[13], and Kemeny's promo line is that a century of progress can condense into a decade ^[9]. Both of those push on the review queue, which is the argument for putting the stop somewhere a schedule cannot negotiate with it.