代理與實體系統
Anthropic Previews MHS, a Unified Driver Layer for AI Agents to Control Laboratory and Manufacturing Equipment
The Model Hardware Standard abstracts device state and operations into discoverable primitives such as `read` and `write`, which can be orchestrated through MCP, a CLI, or programmatic interfaces. It has been used with liquid handlers, robotic arms, and quantum-computer lasers, but its specification and code remain unpublished, while the security evidence comes mainly from partner proofs of concept.
Anthropic has opened a limited research preview of the Model Hardware Standard (MHS), an attempt to create a common control layer that enables AI agents to operate laboratory and manufacturing equipment across different vendors. Each device still requires an MHS driver that translates its existing API, legacy scripts, or even graphical interface into basic operations such as `read` and `write`. The driver also publishes measurable states, configurable parameters, device capabilities, and enforced safety boundaries in a standard format. Agents can control equipment through MCP, the command line, or programmatic interfaces. For high-speed or long-running operations, workflows discovered by an agent can be compiled into deterministic scripts, avoiding the need to wait for model inference at every step.
The technical value of this design extends beyond controlling a single instrument with natural language: it allows agents to discover and orchestrate equipment distributed across different computers and incompatible interfaces. In a proof of concept with Genentech, Anthropic enabled Claude to coordinate liquid handlers, robotic arms, and microplate readers, adjusting pipetting speeds for water and viscous protein samples based on absorbance results. A CMU team said it completed in eight hours an integration that would normally take weeks, and that the system blocked equipment actions under all six artificially induced fault conditions. QuEra separately said an agent reduced the time required to realign a quantum computer’s lasers from several minutes of manual work to a few seconds.
Engineering teams should not yet treat MHS as a validated autonomous experimentation platform. Access is currently application-based, and the full protocol, driver implementations, threat model, and conformance tests have not been open-sourced. The case studies were also selected by Anthropic and its partners, with no independent evaluations comparing different models, measuring long-term failure rates, or testing high-risk equipment. In the Genentech experiment, Claude repeatedly retried after encountering a physical problem caused by bubbles and still required guidance from an expert. Key issues to watch next include tiered permissions, hardware-level guarantees for emergency stops, audit logs for commands and sensor data, and whether an open-source release can prevent vendor-specific drivers from fragmenting into incompatible dialects.