Anthropic Just Gave AI Agents a Body

Yesterday, Anthropic opened a research preview of the Model Hardware Standard (MHS) — a specification for AI agents to operate physical devices. Not generate code for them. Not simulate them. Drive them. Microscopes, liquid handlers, robotic arms, lasers, plate readers. Real machines, real samples, real consequences.

The results from the first four partners — Genentech, QuEra Computing, HHMI Janelia, and the University of Washington — are the most concrete demonstration I've seen of AI crossing the digital-to-physical boundary.

The Playbook Is Familiar

If you've been watching Anthropic, you know the pattern. First came the Model Context Protocol (MCP) — a standard for AI agents to interact with software tools and data sources. Now comes MHS — the same idea, but for hardware.

The problem MHS solves is embarrassingly universal. Every lab and factory floor is a museum of vendor-specific interfaces. Devices from different manufacturers speak different protocols, run different languages, require different SDKs. Integrating a new piece of equipment takes weeks to months, requiring specialists to write bespoke bridge code for every combination.

graph LR
  A[AI Agent] -->|MCP| B[Model Context Protocol]
  B -->|MHS Driver| C[Standardized Primitives
read / write / discover]
  C --> D[Microscope]
  C --> E[Liquid Handler]
  C --> F[Robotic Arm]
  C --> G[Plate Reader]
  C --> H[Laser System]
  style A fill:#a78bfa,stroke:#a78bfa,color:#fff
  style B fill:#52525b,stroke:#52525b,color:#fff
  style C fill:#3b82f6,stroke:#3b82f6,color:#fff

MHS introduces a standardized driver with three simple primitives — read, write, and discover. Every device exposes its state through these. Natural language tags encode safety limits (e.g. "this robot arm weighs 30kg, maximum speed 0.5m/s"). Once onboarded, any agent can operate any device without custom integration. The driver auto-generates a reference file telling the agent what the device can measure, adjust, and what safety boundaries are enforced.

The Numbers That Matter

QuEra Computing: 150s → 6s, 58% → 99.3%

This was the case that made me re-read the press release. QuEra builds neutral-atom quantum computers. Inside those machines, titanium-sapphire lasers must hold their frequency to one part in a trillion. When the lock drifts — from temperature changes, vibration, someone opening a door — an expert needs 5–10 minutes to recover it.

QuEra had already automated this with a bespoke script written by a cross-functional team over months. That script took 150 seconds per attempt and succeeded 58% of the time. The linear sequence it used couldn't absorb mid-process disturbances — if a parameter shifted during recovery, the whole thing had to restart.

They handed the same problem to Claude through MHS. The agent ran overnight, cycling through four roles — one proposing hypotheses, one writing code, one executing against the live laser, one evaluating results. By morning, recovery took 6 seconds and worked 96% of the time. In a blind test across 700 trials: 99.3% success. The hard cases took 10–14 seconds versus 5–10 minutes for a human. The simple ones: under a second.

Claude transformed a linear recovery script into a decision tree. It discovered shortcuts a human couldn't — if the frequency barely moved, most controls don't need touching. No operator would risk skipping checks. Claude ran thousands of disturbance iterations until patterns emerged that no human had the patience or speed to find.

Genentech: AI as a Lab Scientist

Genentech's researchers used MHS to automate the BCA protein assay — a standard procedure requiring coordination across a liquid handler, a robotic arm, and a plate reader. Claude was given the protocol and asked to optimize fluid dynamics for two different liquids: water (simple) and viscous BSA protein solution (foamy, bubble-prone).

Claude ran trial transfers, read the absorbance results, calculated RMSE against expert-performed transfers, and iterated. It converged on 140 µL/s for water (0.016 RMSE) and 10 µL/s for BSA (0.181 RMSE) — parameters the automation experts confirmed as correct.

But the interesting part was the failures. When bubbles formed during mixing — a physical phenomenon Claude couldn't intuit — it defaulted to retrying with different parameters, which only made the bubbles worse. The researchers had to inform Claude that the error was bubbles and it needed to move to a clean well. Once told, it held that context for the rest of the run. This is the current ceiling: models reason broadly but lack embodied physical intuition. The fix is codifying that intuition into reusable "skills" the agent can retrieve next time.

HHMI Janelia: 7 Programs, 1 Click

Virginie Ruetten, a neuroscientist studying sleep in zebrafish, had a rig cobbled together from seven vendor programs running in MATLAB, Python, and C#. Starting an experiment required launching them in a fixed order — get it wrong and the session was lost. Integrating new hardware took days.

MHS replaced all those point-to-point connections with a single shared-memory state dictionary. Starting an experiment is now one click. Adding a new camera took minutes, and its output could immediately feed back into controlling mirrors the camera was imaging.

More impressively, Ruetten wrote a modular online analysis framework — transforms on data streams that work across any device, because every stream is in the same format. An agent now helps identify oscillatory neural populations, zoom in on them, and run closed-loop experiments that would have been impractical to program deterministically.

University of Washington: 8 Hours vs. Multiple Weeks

The Baker and Pinglay labs used MHS to automate a serial dilution curve — the kind of routine assay that costs ~$100 and a week of labor per candidate, repeated for thousands of protein designs. From raw, non-automated equipment to a completed dilution curve with an autonomous rerun: 8 hours. The agent decided the first curve had poor fit (R² < 0.9) and independently reran with a refined concentration range. The second run achieved R² > 0.98.

The Meta: This Is the Inflection Point

MHS is a standardization play, and Anthropic is executing it ruthlessly. MCP for software tools, MHS for hardware. A two-layer protocol stack that lets any agent — not just Claude — operate any tool, on any device, anywhere.

The QuEra result is the one to stare at. Not because 99.3% is perfect (it's not, and the 0.7% still needs a human), but because the improvement came from the agent iterating on its own code overnight. The final output was a deterministic, inspectable script — not a black-box model call. Claude didn't just execute a task; it discovered a better algorithm for the task and compiled it into production-ready code.

The limitations are real. Bubbles. Physics. The things that don't fit in a text token. MHS can't give Claude common sense about fluid dynamics any more than MCP can make it understand your codebase without RAG. But it gives it a standardized bridge, and that changes the cost structure of integration.

Right now, MHS is a research preview. Partners only. But open-sourcing is on the roadmap. When that happens, every lab, every clinic, every factory with programmable equipment will have a choice: spend weeks wiring up custom integrations, or spend hours describing your hardware to an agent in natural language and let it figure out the rest.

I know which one I'd bet on.