The problem
How it works
Anthropic’s video tells how the Model Hardware Standard grew out of work with HHMI Janelia Research Campus, and the announcement explains the design. Each instrument gets a driver with simple read and write commands and can be discovered on the network. Plain-language notes on things the code does not show, such as a robot arm’s weight and the safety limits it enforces, become a reference file for the agent. The agent drives devices through MCP, a command line or code files, sequences steps across instruments, watches results and adjusts settings, and can turn what it learned into a script that runs without reasoning at each step. Partners used it with liquid handlers, plate readers, robot arms, qPCR machines and microscopes.
Results and limitations
Marked reported from Anthropic’s and its partners’ own accounts: Carnegie Mellon researchers say serial-dilution dose-response experiments ran about three times faster with an agent coordinating instruments across three computers, and QuEra says an agent-built controller recovered its quantum computer’s laser lock 99.3% of the time without a person stepping in. No independent evaluation is cited.
Before you try it
MHS is a research preview for selected labs and manufacturers who apply, and Anthropic says it will be open-sourced later. It is model-agnostic and works with any device that has a programmable interface. Physical equipment can ruin samples or hurt people, so set the safety limits in each driver, begin with dry runs and keep someone able to stop the run. Results depend on your setup and were not independently reproduced by Top 200 Repos.
Tools and repositories
No ranked repository is involved. This use case still stands on its own source.
- Tools
- Model Hardware Standard · An MCP-capable agent such as Claude
- Prerequisites
- Acceptance into the MHS research preview and instruments with a programmable interface.
Try it yourself: Describe an instrument for an agent
Fill the fields to adapt the prompt. Empty fields stay as visible [placeholders].
Help me write the plain-language notes for an agent-controlled [Instrument] in our lab. It can measure [Measurements] and adjust [Settings]. Ask me about anything the code would not show, such as [Facts not in the code], then produce a short reference: what it does, the safe range for each setting, what must never happen, and when the agent must stop and call a person.Have the person who runs the instrument every day review the notes before any agent uses them.
Sources
- AI models can now help run physical science experiments (opens in a new tab)
Vendor post · youtube.com · Published 27 Aug 2026 · Source checked 13 Nov 2026, 19:26 UTC
What this source supports
- MHS is a standard for AI agents to safely operate physical equipment in research and manufacturing
- Began as a collaboration between Anthropic’s Beneficial Deployments team and HHMI Janelia Research Campus
- Research preview with select partners
- Previewing the Model Hardware Standard (opens in a new tab)
Vendor post · anthropic.com · Published 27 Aug 2026 · Source checked 13 Nov 2026, 19:26 UTC
What this source supports
- A standard driver with read and write primitives, device discovery, and natural-language tags that generate a reference file with safety limits
- Agents control devices through MCP, a CLI or code files and can chain commands into deterministic scripts
- Carnegie Mellon ran serial dilution dose-response experiments about three times faster
- QuEra’s agent-developed controller recovers the laser lock 99.3% of the time without human intervention
- Model-agnostic, works with any programmable device; research preview by application ahead of open-sourcing