Shared-Memory AI Cuts Lab Setup From Months to Days
A Janelia researcher used Claude Code to build a shared-memory system that reduces lab instrument integration from months to days, spawning a new open standard called MHS.
Arco Bast, a postdoctoral researcher at Howard Hughes Medical Institute's Janelia Research Campus, used Claude Code, Anthropic's AI coding tool, to solve a problem that had cost him months of work: getting cameras, scanners and sensors from different manufacturers to cooperate inside a single microscope setup. His solution, a shared-memory software layer, compressed that integration work from months into days. He and his Janelia colleagues then partnered with Anthropic to formalise the approach as the Model Hardware Standard, or MHS, aiming to do for lab instruments what USB did for peripherals.
What happened
| Detail | Fact |
|---|---|
| Researcher | Arco Bast, postdoctoral researcher, Janelia Research Campus (HHMI) |
| Tool used | Claude Code (Anthropic) |
| Time saved | Months of integration work reduced to days |
| New standard | Model Hardware Standard (MHS), developed with Anthropic |
| Analogy cited | USB: one protocol connecting many device types |
Bast’s microscope can image thousands of neurons simultaneously, but its components each had separate software stacks and private memory spaces. Any time he wanted to run an experiment differently, information had to travel between isolated programs via the operating system, and Bast often spent months engineering those connections before doing any actual science.
His insight was architectural: instead of routing data through a central OS, what if every component read and wrote to a single shared memory pool? He built that layer using Claude Code, and the result was an always-current picture of what the experiment was doing, visible to every instrument at once.
What is the Model Hardware Standard?
After validating the approach internally, Bast and his Janelia colleagues worked with Anthropic to develop it into MHS, a proposed common protocol for AI systems interacting with physical scientific equipment and manufacturing hardware. The researchers compare it to USB: if instrument makers adopt MHS from the start, their hardware would be interoperable without custom glue code.
The practical upside for scientists: instead of writing large amounts of bespoke integration software, researchers could potentially connect equipment from different manufacturers and issue instructions in plain language. Think “scan this region at 2x speed if the signal drops below threshold” rather than months of low-level programming.
Why does the shared-memory approach matter for AI agents?
The architecture opens a second door beyond faster setup. Because every component updates a shared data structure in real time, an AI agent can watch the experiment as it runs, spot patterns, and adjust equipment on the fly. That is a meaningful shift from today’s typical use of AI in science, which mostly happens after data collection is complete.
Researchers at Janelia believe this could extend to hypothesis generation, experimental design, and mid-experiment modification. In neuroscience specifically, they point to potential applications in studying memory formation and disorders such as Alzheimer’s disease. Those are ambitious goals and still speculative at this stage, but the underlying plumbing is real and working today in Bast’s lab.
Why it matters
Most coverage of AI in science focuses on model capabilities: better protein folding, faster drug screening, smarter image analysis. This story is about something more foundational: the interface between AI software and physical instruments. If MHS gains adoption among hardware manufacturers, it could lower the barrier to AI-assisted experimentation across biology, chemistry, materials science and beyond.
For businesses thinking about AI integration into their own operational workflows, the parallel is instructive. The bottleneck is rarely the AI model itself. It is usually the plumbing: getting disparate systems to share state so the AI has something coherent to act on. Bast’s shared-memory insight applies equally to enterprise software stacks where CRMs, analytics tools and automation platforms each sit in their own silo.
Our take
The MHS concept is genuinely interesting and the underlying engineering problem is real. Scientists spending months on equipment wiring before they can ask a single scientific question is a genuine waste, and a shared-memory architecture is a clean solution to a coordination problem that every complex system faces.
That said, the USB comparison should be treated with some skepticism. USB succeeded partly because a small number of large hardware companies agreed to back it early. Getting lab instrument manufacturers, who often compete partly on proprietary software ecosystems, to adopt a common AI-facing standard is a harder organisational problem than solving the technical one. Watch for early adopter announcements from named instrument makers before assuming broad uptake.
Claude Code as a development accelerator is the more immediately verifiable claim here. Bast used it to compress months of integration work into days, and that pattern, AI-assisted plumbing work cutting integration time dramatically, is consistent with what we see in client projects we ship. The real question is whether the output is maintainable long-term, not just fast to produce.
If you are running complex multi-tool workflows and spending more time on integration than on the actual work, this story is a prompt to revisit your own architecture. A shared state layer, whether for lab instruments or for automated business workflows, is often the missing piece.
What to do about it
- Follow the MHS specification as it becomes public. If you build or procure hardware-adjacent software, understanding the standard early puts you ahead of mandatory retrofits later.
- Audit your own integration stack. List every system that currently maintains its own data silo and map where information has to travel through a central broker. That map tells you where you are losing time.
- Test Claude Code or similar AI coding tools on a real, bounded integration task. Bast’s result is the most credible benchmark you have right now: months to days on a well-defined plumbing problem.
- Stay cautious on the “AI as active experiment participant” framing. The real-time adjustment capability is promising but requires careful validation before any process where errors have consequences.
The practical takeaway: your biggest AI bottleneck is probably not the model, it is the data plumbing around it.
Frequently asked questions
What is the Model Hardware Standard (MHS)?
MHS is a proposed common protocol, developed by Arco Bast, his Janelia Research Campus colleagues and Anthropic, that lets AI systems safely interact with physical scientific equipment and manufacturing hardware. The goal is to make multi-vendor lab instruments interoperable without custom integration code, similar to how USB standardised peripheral connections.
How did Claude Code help with lab instrument integration?
Arco Bast used Claude Code, Anthropic's AI coding tool, to build a shared-memory software layer that lets cameras, scanners and sensors in his microscope setup read and write to a single common data pool. Work that previously took months was completed in days.
What is shared-memory architecture in the context of lab equipment?
Normally, each instrument component maintains its own private memory and exchanges data through a central operating system, which is slow and requires custom plumbing. A shared-memory approach gives every component direct access to the same continuously updated data pool, removing that bottleneck.
Can AI agents control lab experiments in real time?
The shared-memory architecture makes this possible: an AI agent can monitor live experiment data and adjust equipment on the fly rather than only analysing results after the fact. Bast and his Janelia colleagues cite this as one of the key benefits of MHS, though broad deployment is still in early stages.
