Safe closed-loop agentic dose-response experiments using MHS

During the limited Model Hardware Standard (MHS) research preview (announced by Anthropic), I led Carnegie Mellon’s work using MHS. The preview specification is provisional and model-agnostic; we used Claude as one agent.

An agent operated a liquid handler, plate reader, robotic arm, and cameras across three computers with incompatible interfaces. Serial-dilution dose-response experiments ran about three times faster than our prior workflow. The work is in the MHS announcement and in Nature (24 September 2026); CMU SCS also posted on LinkedIn and X.

I wrote the MHS drivers for all three instruments and the orchestration layer. The agent ran the protocol, rejected a saturated curve, and reran with a lower maximum concentration (\(R^2 > 0.98\)), within device-declared safety bounds. Integration took about eight hours, compared with weeks of per-instrument integration.

Full write-up, photos, and video: Safe closed-loop agentic dose-response experiments using the Model Hardware Standard.