The load-bearing artefact in this design is not the transport layer, it is a text file. Anthropic says the MHS driver carries tags in which the operator writes, in plain language, what code cannot show, and its own example is the weight of a robot arm, which matters for knowing how to manipulate it safely and has historically sat in a paper manual or in someone's head [15][10]. From those tags the driver generates a reference file stating what the device can measure, what can be adjusted, and what safety limits will be enforced [11]. Whoever answers that setup interview is authoring the operating envelope an agent will treat as ground truth.
Size the compression at its endpoints before accepting it. Two months of one specialist at 40 hours a week is about 320 hours, so against a one-hour setup the ratio is roughly 320 to 1, and against a ten-minute one closer to 1,900 to 1 [13]. Those endpoints are Anthropic's own framing rather than a measured benchmark, and the preview post gives no per-device timings [14]. Even the conservative end is enough that bespoke integration work, which the company says currently requires specialists because most devices do not talk to each other, stops being the reason a facility runs one instrument at a time [4].
A facility manager's objection is fair: the drivers were never the expensive part, qualified hands and physical interlocks were, and a shared specification buys neither. That is right about cost and beside the point on sequencing, because removing the cheap constraint pushes throughput against the expensive one, and the expensive one here is a named person willing to authorize work that continues after everyone has gone home.
The supervision question gets sharper exactly where the standard is most useful. Anthropic describes agents sequencing steps across instruments, adjusting parameters as conditions change, and in some cases recovering from hardware errors without intervention [6]. It also says that for long-running tasks, or when devices must be driven faster than the agent's online reasoning allows, the agent chains driver commands into code files [16]. In that mode the per-step reasoning a reviewer would want to inspect is not happening per step, and commands can arrive by three routes at once: MCP, a command line interface, and API code files [12]. Anyone assigning responsibility for an overnight run needs to know which route issued the instruction that ruined the sample.
The board-deck version reads cleanly: a model-agnostic specification that works with any programmable device and is reachable through standard protocols such as the Model Context Protocol [8], provenance from a collaboration with HHMI Janelia Research Campus [3], and integration measured in minutes instead of months. What it omits is that the post names no other preview participants, sets no open-source date, and does not say who is accountable when an agent operates equipment unattended [14]. This quarter the decision in front of a lab director is narrow, which instruments to expose and to whom. Next quarter's consequence is that reference files written quickly during a preview become the de facto safety limits for as long as they keep working, because configuration that works is rarely reopened.
