Most technology standards begin in a committee. This one began with a neuroscientist who could not get his own microscope to talk to itself.
Anthropic’s Model Hardware Standard, previewed on 27 August 2026, has a traceable origin that is unusually specific and unusually human: Arco Bast, a postdoctoral researcher in the Spruston lab at HHMI’s Janelia Research Campus, built a shared-memory approach so the components of his imaging rig could coordinate — and that prototype became the basis of the standard. The MHS origin story is worth telling accurately, because it explains what the standard is actually for far better than any architecture diagram does.
The Problem: One Instrument, Several Languages
A modern imaging rig is not one machine. It is a collection of machines from different vendors that happen to be bolted to the same optical table: cameras, lasers, mirrors, motorised stages, electrophysiology hardware. Each arrives with its own control software, and — this is the part that surprises people outside the field — they frequently run in different programming languages.
Anthropic’s own description of the Janelia setup puts the problem plainly: the components come with vendor control software and no common interface, and “they often run in different programming languages, too: the detectors run in MATLAB, the cameras in Python, the electrophysiology in C#.”
The consequence is the sentence that makes the whole problem click:
“A quantity held by one program, such as the position of the stage, is therefore unknown to the others.”
That is the entire difficulty in one line. The stage knows where it is. The camera does not. If you want the camera to do something because the stage is at a particular position, you have to write a bespoke connection between two programs in two languages — and then another one for every other pair that needs to coordinate. The number of connections grows with the square of the number of instruments, and each one is a small custom translator that someone has to write and maintain.
Bast’s own account of the frustration, as reported in HHMI’s coverage, is measured in a way that will be familiar to anyone who has fought their tools: the mesoscope, he says, “is an amazing microscope that, in principle, would allow me to do the many things that are well suited to measuring what the cells are doing, but not with the control it had.”
(A sourcing note, because it matters: hhmi.org returns HTTP 403 to automated retrieval, so Bast’s quotes here reach us through outlets that closely republish HHMI’s own release rather than from the HHMI page directly. Treat them as corroborated through syndication rather than independently primary-verified.)
The Idea: Stop Connecting Programs to Each Other
Bast’s insight was to change the topology rather than improve the translators.
His description of it: “if the different components could instead share a pool of memory that each could read and write to directly, they could work together.”
That is the whole design. Instead of n×n point-to-point connections between programs, every device writes its state into one shared dictionary living in shared memory — a region the operating system lets many programs access simultaneously. The stage writes its position there. The camera reads it. Neither needs to know the other exists, what language it is written in, or how to speak to it.
Anthropic’s announcement describes the resulting architecture, in the words of Virginie Ruetten — a scientist in the Ahrens lab at Janelia whose own rig uses it — as replacing “those point-to-point connections with a single interface,” where each device is “described and onboarded once, and its variables, controls, and sensor values are recorded in a single dictionary that lives in shared memory.”
Anyone who has worked on integration problems will recognise the move immediately. It is the same shape as a message bus, a blackboard architecture, or a shared database: replace many bilateral agreements with one common medium. The reason it is worth noticing here is not novelty. It is that a working scientist arrived at it because his experiment demanded it, and then it turned out to generalise.
For his own work, Bast reports the integration effort dropping from months to days. (That is an N=1 self-report about one researcher’s own experiment, not a measured benchmark — and it is worth keeping it labelled that way, because the number travels well and its scope does not.)
How It Became a Standard
The path from a postdoc’s workaround to a published specification ran through one collaboration.
Alek Kemeny, on Anthropic’s Beneficial Deployments team — a division led by Elizabeth Kelly — worked with Bast, and the shared-memory pattern became a company initiative rather than a lab tool. That is the entire bridge: a researcher with a specific problem, and someone inside a company who recognised the pattern was general.
(Worth flagging for anyone researching this further: a tidy line attributed to Kemeny — that what MCP did for software, MHS will do for hardware — circulates widely but does not appear in Anthropic’s announcement, where he is named as a contributor and is not quoted at all. It is a good example of a quotable sentence acquiring a source through repetition.)
Elizabeth Kelly has been notably direct about the dual motive, telling CNBC: “We built this for science to sort of show the promise of AI, but there’s also huge benefits here for enterprise and for industry.” That candour is worth reading as candour. A company describing both the sympathetic and the commercial reason for a project, unprompted and in the same breath, is being unusually straightforward — not caught out.
Why the Origin Explains the Design
Knowing where MHS came from makes several of its choices legible in a way the announcement alone does not.
Why natural-language tags? Because the person who knows a laser’s safe power ceiling is the scientist who uses it, not a software engineer writing a schema. Bast’s problem was not that his instruments lacked documentation — it was that the documentation lived in manuals, vendor software and his own head. A tagging step that a researcher can complete in prose, or by answering an agent’s questions, is a solution shaped by someone who has actually had to onboard an instrument.
Why a shared dictionary rather than a protocol? Because the original problem was not remote control, it was mutual awareness between programs on one rig. Shared memory is the natural answer to that specific problem, and it is a notably different starting point from OPC UA or SiLA 2, which began from cross-vendor networked interoperability.
Why “onboarded once”? Because the pain being solved is repeated integration work — the n² translators — not any single connection.
This also explains a limitation honestly. A design that grew from making one rig’s components coordinate is not automatically a design for authenticating unknown devices across a network, or for enforcing safety limits against an adversary. Those are different problems, and MHS’s public material, as of 29 August 2026, does not yet address them — which is unsurprising given where it started.
The Part of This Story Worth Keeping
Strip out the standard entirely and something durable remains, which is why this story is worth telling even to readers who will never touch a mesoscope.
A working scientist solved his own tooling problem, and the solution generalised. Not a roadmap item, not a strategic initiative — a person who could not run the experiment he wanted, building the thing that would let him run it. HHMI’s own framing of the story is “How One Postdoc’s Problem Solving is Changing the Way Scientists Work,” and that is the right emphasis.
It is also a reminder about where infrastructure comes from. The instinct is to assume that standards descend from committees to users. Frequently the direction is the opposite: someone with a concrete problem builds something specific, it works, and the generalisation happens afterwards — often by a different set of people, in a different institution, with different incentives. Both halves of that sentence matter. The origin is a scientist. The standard is a company’s.
What that means practically is that the origin story is a good guide to what MHS is good at — reducing integration friction between instruments that were never designed to cooperate — and a poor guide to what it has not yet been asked to do. The Janelia rig had no adversary. Most labs do not. But a specification distributed to the world eventually meets one, and that is a different design problem than the one Bast set out to solve.
Frequently Asked Questions
Who created the Model Hardware Standard?
It originated with Arco Bast, a postdoctoral researcher in the Spruston lab at HHMI’s Janelia Research Campus, who designed a shared-memory approach so that the components of his imaging rig — lasers, mirrors, cameras — could coordinate across different programming languages. Alek Kemeny, on Anthropic’s Beneficial Deployments team, worked with Bast to turn that prototype into the basis of MHS, which Anthropic previewed on 27 August 2026.
What problem was Bast actually trying to solve?
Instruments on one optical rig came from different vendors, each with its own control software, frequently written in different languages — detectors in MATLAB, cameras in Python, electrophysiology in C#. As Anthropic’s description puts it, “a quantity held by one program, such as the position of the stage, is therefore unknown to the others.” Making any two coordinate meant writing a custom translator, and the number of translators grows with the square of the number of instruments.
What is the shared-memory dictionary?
Instead of connecting programs to each other, every device writes its state into a single dictionary held in shared memory — a region of the computer’s memory the operating system lets many programs access. Each device is described and onboarded once, and its variables, controls and sensor values are recorded there. The camera can read the stage’s position without either program knowing anything about the other.
Did MHS really cut integration time from months to days?
That figure comes from Bast, about his own experiment, and should be read as an N=1 self-report rather than a measured benchmark. A separate and narrower figure appears in Anthropic’s announcement from Virginie Ruetten of Janelia’s Ahrens lab, who describes adding a new camera — previously “a multi-day project” — taking “only a few minutes.” The two claims have different scopes and should not be merged.
Does the origin story tell you anything about MHS’s limitations?
Yes, usefully. A design that grew from making one rig’s components mutually aware is well shaped for reducing integration friction between cooperating instruments. It is not, by construction, a design for authenticating unknown devices across a network or for enforcing limits against an adversary — the original rig had no adversary. Those are genuinely different problems, and MHS’s public material, as of 29 August 2026, does not yet address them.
📦 Recommended Resources
HHMI’s own account of Bast’s work and Anthropic’s announcement are both free and short, and they are the primary sources for this story. These books are for the wider theme it illustrates: how tools shape what science can ask, and what happens when the instruments start running themselves.
- Reproducibility and Replicability in Science — the National Academies consensus study, and the natural companion to this story. Bast’s problem was that his instrument could not do what his experiment needed; the flip side, once instruments can run themselves overnight and adjust their own parameters, is what the scientific record of that run actually contains. This report is the most rigorous treatment of that question, and its distinction between reproducing an analysis and independently replicating a finding is exactly the right lens. Free to read online from the National Academies.
- Laboratory Automation in the Chemical Industries edited by David G. Cork and Tohru Sugawara — the book-length history of the integration problem this article describes, from a field that has been fighting it since long before AI agents existed. Old (2002) and expensive, and worth it only if lab integration is genuinely your world — but its age is part of the point: the n²-translators problem Bast ran into is decades old, and reading someone describe it in 2002 is a useful corrective to the sense that any of this is new. (Verified against bibliographic record; Amazon’s listing has been intermittently unreachable.)
- Normal Accidents: Living with High-Risk Technologies by Charles Perrow — for the closing thought of this article. Perrow’s subject is what happens in tightly-coupled systems where components interact in ways nobody designed. A shared dictionary that lets every instrument read every other instrument’s state is, by construction, an increase in coupling — which is a real gain in capability and a real change in failure modes. Worth reading before you conclude that more integration is straightforwardly better.
✅ Pro tip: The generalisable lesson here costs nothing: when two systems will not cooperate, consider changing the topology rather than improving the translators. Replacing n² bilateral connections with one shared medium is the move behind message buses, blackboard architectures and shared databases — and it was available to a neuroscientist who had never heard those terms.