Experiment-as-Code Labs: A Declarative Stack for AI-Driven Scientific Discovery
AI agents can now design and execute real-world lab experiments via declarative configurations.
A new paradigm called Experiment-as-Code (EaC) Labs aims to untether AI agents from purely digital environments, enabling them to control real-world lab instruments. Proposed by Zhenning Yang, Yuhan Chen, and colleagues from academia (including Venkat Viswanathan and Danai Koutra), the system encodes experiments as declarative configurations that are compiled down to device-level APIs. This allows AI agents to generate hypotheses and experiments in a configuration language, while a systems layer performs program analysis, safety checks, resource assignment, and job orchestration before actuating lab hardware.
The approach is science-, lab-, and instrument-independent, representing a novel synthesis across physical, systems, and intelligence layers. By bridging the gap between increasingly powerful AI agents (like LLMs) and automated lab equipment (which expose programmable APIs), EaC Labs could enable real-time course changes during experiments—e.g., when an AI notices unexpected clues. This contrasts with current autonomous labs that require custom integration. The stack promises to scale AI-driven discovery by letting agents explore physical phenomena as easily as they analyze data, with built-in safety and orchestration.
- EaC Labs lets AI agents encode experiments as declarative configs, compiled to device-level APIs for physical lab control.
- The systems layer includes program analysis, safety checks, resource assignment, and job orchestration to manage real-world risks.
- Stack is science-, lab-, and instrument-independent, enabling any automated lab to integrate with any AI agent.
Why It Matters
Bridges AI reasoning with physical experimentation, unlocking AI-driven breakthroughs in chemistry, biology, and materials science.