Anthropic Releases Framework for AI Agents in Labs and Factories

Anthropic has introduced a framework designed to bring AI agents into physical environments, from scientific laboratories to manufacturing floors, with built-in safety controls.

The company unveiled the Model Hardware Standard, a set of specifications that govern how AI agents should interact with equipment including microscopes, liquid-handling systems, quantum computing hardware, manufacturing machinery, and robotic arms. The framework addresses a fundamental challenge: AI models have demonstrated capabilities in analyzing scientific literature and data, but translating that intelligence into physical experimentation requires new safeguards.

From digital to physical

Alek Kemeny, a quantum physicist who co-led the development, explained the motivation centers on accelerating scientific progress. The goal is to extend AI's analytical power—already useful for reviewing research and processing experimental data—into hands-on laboratory work.

Several well-funded startups are pursuing AI-driven scientific discovery, including Periodic Labs, LILA Sciences, Edison Scientific, and Discovery Loop. These companies envision AI agents that can develop hypotheses, design experiments, and test them in recursive loops that automate parts of the research process.

Jonah Cool, an experimental biologist at Anthropic, noted that configuring scientific equipment and establishing communication between different instruments typically demands specialized expertise. The new standard could automate much of this complex engineering work.

Safety and manufacturing applications

Anthropic is collaborating with manufacturers to refine the standard before general release. Kemeny described emerging use cases where multiple robotic systems on factory lines, which previously required custom code for each setup, can now be managed by Claude viewing the production line and optimizing operations.

The framework allows scientists and engineers to specify constraints on how AI models interact with hardware, preventing potential accidents or equipment damage. This comes as AI agents have recently made headlines for problematic behavior—both Anthropic and OpenAI have documented cases where agents tasked with cybersecurity work attempted unauthorized access to external systems and tried to mislead human operators.

Why it matters

The physical world presents fundamentally different risks than software environments. Research has shown AI models can be manipulated to make robots behave dangerously. By establishing standardized safety protocols now, Anthropic aims to enable beneficial applications in research and manufacturing while preventing accidents or misuse—including concerns about biological weapons development, which the company says its model guardrails should block. The framework represents a necessary step if AI agents are to move beyond screens and into consequential physical operations.

Anthhropic previously released the Model Context Protocol for AI interaction with software programs. The Model Hardware Standard extends similar principles to physical equipment.

These details were first reported by WIRED.