Self-Driving Microscopy with Agents: Progress and Prospects

Agentic AI, powered by large language models, has increasingly been applied within self-driving laboratories and here we review its emerging role in self-driving microscopy. Microscopy is an unusually demanding target as each experiment is a sequence of interdependent decisions about where to measure, at what resolution, through which modality, and often on ambiguous evidence. We review recent demonstrations across atomic force, scanning tunneling, and transmission electron microscopy and examine how agentic systems can lower the expertise barrier to advanced characterization, transfer strategies between platforms, mine laboratory records for reusable protocols, generate and screen hypotheses, and bring simulation into the experimental loop. We also highlight and examine the limitations that currently constrain these systems. Agents can lose track of the scientific goal as a session outgrows their context window. They also remain poor judges of their own competence, making failure particularly difficult to anticipate in open-ended research. Flexible reasoning must therefore sit on deterministic, physically constrained control, with safety limits enforced outside the model. Human oversight can be the safeguard, but a human acting only as an approval or rejection gate risks forfeiting what scientists contribute most — questioning how the problem was framed in the first place. Designing agentic systems that preserve that contribution and address the current limitations is, in our view, the defining challenge for the field.