What happens when an agent that learned to retry flaky APIs gets a pipette? At Genentech, Claude recovered from tip pickup failures on its own. Then the sample foamed. It kept retrying the same well.
Name the failure: software instinct on wetware
Call it software instinct on wetware. In the software world that agents grew up in, a failed tool call is often safe to repeat. Timeouts, 500s, empty JSON: try again. In a protein assay, foam is not a timeout. It is surface tension, gas, and agitation. Retrying the same well made the foam worse until a human steered the agent toward a clean well and gentler mixing.
That vignette is not a cute lab anecdote. It is the gap Anthropic opened a product-shaped standard around on 27 August 2026.
MCP taught discovery. MHS puts discovery on hardware.
Anthropic's research preview of the Model Hardware Standard (MHS) is a shared driver and discoverability layer so agents can operate programmable lab and factory devices: microscopes, liquid handlers, robotic arms, lasers, even quantum-control hardware. The primitives are almost insultingly small: read (get temperature) and write (set temperature). Devices advertise themselves in a common format. Operators tag physical facts that code cannot see (arm mass, safe laser ranges) in natural language. The driver compiles those tags into a reference file with enforced safety limits.
Control rides three rails at once: MCP, CLI, and code files. The claim is model-agnostic. The published case studies used Claude because Anthropic runs the preview. The architectural point is sharper than brand: MCP standardized how agents find software tools. MHS tries to do the same one layer down, where a wrong write can ruin a sample or misalign a laser.
Integration time is the commercial pitch. Carnegie Mellon wired a robotic arm (job-file scheduler, no normal API), a CyBio Felix liquid handler (legacy ActiveX/COM), and a Varioskan LUX plate reader that was GUI only in about eight hours, including an autonomous dilution-curve rerun. Typical vendor timelines are measured in weeks. University of Washington Baker and Pinglay labs connected six instruments in under a week. Janelia compressed an imaging run from weeks toward a day.
The numbers that survived a blind test
Genentech's BCA assay is the closed-loop story. Claude coordinated liquid handler, arm, and plate reader, then tuned flow rates against expert ground truth: water near 140 uL/s (RMSE 0.016), viscous BSA near 10 uL/s (RMSE 0.181). Experts called the parameters reasonable. Tip failures recovered without a human. Foam did not.
QuEra is the hard-physics receipt. Laser stabilization on their quantum hardware moved from roughly 58% success with slow recovery to a development regime near six-second recovery at high success, then a blind test at 99.3% (695 of 700). That is not a chat demo. It is a control loop that either holds the beam or it does not.
CMU's dose-response work claimed roughly 3x faster serial dilutions than manual runs, concentration curves with R-squared above 0.98, and mid-experiment range adjustment when early points said the original window was wrong. Safety limits still sat in the driver layer, not in the model's "judgment."
Explore, then freeze the procedure
Anthropic's own description of Claude on hardware is the second half of the story. The agent does not stay in free-form tool chat for the whole experiment. It explores like a scientist: adjust laser, watch camera, repeat. When the sequence stabilizes, it packages the learning into a deterministic code file so the device path runs as one command without per-step reasoning.
That pattern rhymes with the software stack the industry already built. Long-horizon agents write skills, scripts, and harnesses because weights alone do not hold procedure. MHS makes the same move in atoms: reason while uncertain, compile when stable, enforce limits in the driver rather than in the next token.
It also explains why the foam failure is structural. A compiled procedure inherits whatever physical model the exploration phase got wrong. If the exploration phase treats foam as a flaky read, the frozen script will thrash the same well faster.
Why it is still a research preview
Anthropic is not open-sourcing MHS yet. The stated reason is not marketing polish. It is a physical safety roadmap: misuse risk, policy, and evaluation for agents that can move real mass and energy. Vendors already lining up support include AWS Strands Robots, Universal Robots, Tecan, QIAGEN, Danaher, Doosan, Automata, Hugging Face LeRobot, MBF Bioscience, and even Raspberry Pi class hardware. That breadth is the USB analogy talking: standardize the cable before every bench rebuilds a custom bus.
USB peripherals rarely foam. Lab samples do. The honest limit in Anthropic's own write-up is that models still learn the physical world from text and images, not from contact. MHS shortens the path from intent to instrument. It does not give the model a body. Expert oversight remains part of the design, not a temporary embarrassment.
What the foam is telling the rest of the stack
If you never touch a liquid handler, the lesson still lands. Agent platforms spent 2025-2026 teaching models that tools are retryable, discoverable, and cheap to re-call. That was correct for HTTP and SQL. It is actively wrong for irreversible physical state.
The next product fights will not only be about longer context or cheaper tokens. They will be about which layers refuse a write: the model, the skill, the driver safety file, or a human at the bench. MHS is Anthropic planting a flag that the refusal belongs in a shared hardware layer, the same way MCP planted a flag for software tools.
Watch three things next: whether non-Claude harnesses actually clear the research-preview gate; whether safety evaluations publish failure taxonomies beyond the foam anecdote; and whether "compile the successful loop" becomes the default pattern for physical agents the way skill files became the default for software agents. The pipette already answered the first question about blind retry. The industry has not finished answering who is allowed to write.