Hardware was always invited

Anthropic just announced the Model Hardware Standard — a specification for AI agents operating lab equipment. Read/write primitives, device discovery, natural-language capability tags with enforced safety limits, MCP as one of the control surfaces, chained commands for long-running tasks.

We read it with the particular pleasure of recognizing a shape.

The shape

Mycelium's README has said, publicly since July 9:

A memory and coordination layer for anything that can make an HTTP request. A camera, sensor, or actuator? POST readings, subscribe to commands, or claim jobs — hardware is a first-class citizen.

And the machinery behind that sentence is older than the sentence. The public history of SoftBacon-Software/mycelium opens on March 1, 2026 already mid-stride — executing step-sequenced plans (commits reference "Plan #9, Steps 81–91"), with approvals, and with a drone system: remote machines that claim dispatched jobs, execute, and report back, where a step cannot fire before the step it depends on has finished. By March 9 we were fixing drone-dispatch bugs in public — stale jobs, project scoping, worker resilience. The ordinary bugs of a thing that exists.

That is the same anatomy MHS describes, one abstraction level up: discovery is registration, read/write primitives are what a sensor or actuator does when it POSTs readings and subscribes to commands, and "chained driver commands for long-running tasks" is plans-and-steps sequencing across machines.

The safety rule we'd underline

One design decision in MHS deserves particular attention, because we arrived at it independently and believe it is the load-bearing one: capability tags in natural language, limits enforced below the model.

Our internal design note from July 24, written for our first physical tenant (a sensed, auto-dosed hydroponic grow system), states it bluntly:

Physical safety lives in hardware/firmware interlocks, never in model judgment. Dose ceilings, max run-time per pump, float-switch cutoffs, fail-closed valves — enforced below the agent. The model may request a dose; the hardware decides what is survivable.

This is the same law our software gates already live by: the compile gate runs in the bridge, not the model, so a clever agent cannot talk its way past it. A description exists for the model; a limit exists in the layer that executes. Any hardware standard that keeps that separation is one we can build on.

Convergence is the strongest validation there is

We don't think anyone took anything from anyone. Ideas whose time has come arrive at multiple doors — we've watched it happen to us before, at three-days-and-71-days resolution, and the lesson was the same both times: being early only counts if you ship, in public, with dates.

So, plainly: the horizontal version of this — hardware as one more citizen of a sovereign, local-first coordination layer, beside agents and humans, on owned machines, with the internet off — has been our stated position for months, and it runs today. MHS is the vertical for laboratories, launched with the partner list a vertical deserves. We welcome it, we intend to implement the standard when it opens, and the fact that the largest AI lab in the world now agrees that models should meet hardware through a standard layer with enforced limits is the best evidence yet that the layer was the right thing to build.

The microgreens don't care who standardized the dosing pump. They grow or they don't. That's the gate we answer to.