Researchers at the National Institute of Standards and Technology have proposed a framework for using teams of artificial intelligence agents to coordinate instruments, experiments, and resources across autonomous materials laboratories.
The peer-reviewed perspective, published in Communications Materials, examines how laboratories could expand beyond today’s narrowly focused self-driving systems. Rather than controlling one experimental workflow or a small collection of instruments, future AI systems could oversee broader research campaigns involving multiple tools, data sources, and scientific objectives.
The authors, A. Gilad Kusne and Austin McDannald, do not describe a fully autonomous materials laboratory already in operation. Instead, they draw on existing autonomous experimentation systems to outline potential architectures and infrastructure needed for broader, lab-wide AI coordination.
Moving beyond individual self-driving systems
Existing autonomous experimentation systems typically use AI to analyze collected data, predict the results of possible experiments, and select the next experiment in a closed loop. These systems have supported focused applications such as materials optimization, phase mapping, and instrument control.
Most remain customized for narrow research campaigns involving a limited number of tools, however. The researchers distinguish these systems from a fully autonomous laboratory, which would contain diverse instruments and support multiple, complex research campaigns under AI coordination.
To reach that level of coordination, the paper presents two possible models. In one, a central AI agent manages laboratory tools through distributed computing modules connected to individual instruments. In the other, a hierarchy of specialized agents manages different tools, tasks, and levels of laboratory operations.
Under the multi-agent model, a lower-level agent could optimize an individual instrument, repeat a measurement after an interruption, or adjust settings when it detects poor-quality data. Higher-level agents could assign research tasks, coordinate experiments across instruments, redirect work when equipment fails, and balance competing demands for laboratory resources.
Managing instruments, data, and schedules
Laboratory-wide AI would require more than automated instruments. The framework would need to schedule serial and parallel experiments, manage sample movement, resolve conflicts over instrument access, and account for differences in run times and resource requirements.
Data integration presents another barrier. Instruments can produce data with different formats, uncertainties, background signals, and systematic variations. The authors emphasize the need for detailed metadata, methods for combining diverse data streams, and standards that allow software and equipment from different vendors to communicate.
The choice of AI also matters. Large language models can support user interactions, but their limited interpretability, computational demands, and potential to produce inaccurate outputs make them risky choices for direct equipment control. The authors suggest that laboratories could use language models for communication while assigning safety-bounded instrument control to more interpretable, physics-informed AI systems.
Testing AI strategies before deployment
Before allowing AI agents to control expensive or hazardous laboratory equipment, the researchers recommend evaluating them in digital or physical sandboxes.
A digital sandbox could use simulations or digital twins to represent instruments, materials, sample handling, and laboratory resources. A physical sandbox could provide access to limited equipment or a surrogate laboratory where managers could test how an AI strategy responds to instrument failures, unusual data, and changing resource constraints.
Advanced Lab Management Certificate
The Advanced Lab Management certificate is more than training—it’s a professional advantage.
Gain critical skills and IACET-approved CEUs that make a measurable difference.
Human oversight would remain necessary. The paper calls for interpretable interfaces, user control at multiple levels, and guardrails that restrict instrument movements, experimental recipes, data sharing, and other potentially unsafe actions. The authors also note that individually safe experimental parameters could create hazardous conditions when combined, making safety evaluation particularly difficult during exploratory research.
For lab managers, the framework identifies the operational foundations that increasingly autonomous laboratories would require: interoperable equipment, standardized data, reliable metadata, cybersecurity and intellectual property controls, defined human approval points, and methods for testing AI behavior before it reaches active instruments. The paper positions multi-agent AI as a potential laboratory management architecture, but substantial technical and governance work remains before such systems can coordinate complete materials laboratories.
This article was created with the assistance of Generative AI and has undergone editorial review before publishing.
