Lab lessons from Anthropic MHS: keep fast control out of the model

Empty-Abalone-2952 · reddit · 2026-09-04

Drawing on Anthropic's MHS preview (a shared driver format for lab and manufacturing equipment, accessible via MCP, CLI, or code), the author explores where agent control of physical hardware breaks down: Claude treated a physical issue as a software error and retried, creating more bubbles; researchers later saved the correction as a reusable skill. In a laser example, Claude ultimately wrote deterministic alignment code so fast control steps didn't depend on model reasoning. The author also untangles failover semantics (channel switching vs. swapping models mid-run) and recommends freezing model choice once a physical run starts, pausing on provider failure, and keeping fast control in deterministic code — MHS standardizes the device interface, but the agent system must enforce the boundary.

Original post →

More from coding & agent

coding & agent channel →