Running one of the world’s most powerful electron microscopes used to take weeks of training. At Lawrence Berkeley National Laboratory, researchers can now describe an experiment in plain English and let an AI agent run it.The system, called TEM Agent, was built at the Molecular Foundry, a U.S. Department of Energy user facility at Berkeley Lab. It controls a transmission electron microscope, an instrument that shows where atoms sit in a material, which determines how strong the material is, how it conducts electricity, how a battery stores energy and how a drug binds to a cell.A microscope that outpaced its usersThe Molecular Foundry’s microscopes can image structures roughly half the size of a hydrogen atom, and scientists from universities and companies travel there to use them. The learning curve is steep. Some visiting researchers arrive with a semester of electron microscopy coursework and still need weeks of hands-on training. Even experts struggle with the data: one detector can produce about 700 gigabytes in 15 seconds, far too fast to process by hand.The result is a bottleneck. The equipment keeps getting more powerful, but the number of people who can run and analyze it expertly can’t keep up.From factory line to conversationThe team already had automated routines for repetitive jobs, such as imaging large numbers of nanoparticles. But those worked like a factory line, said Peter Ercius, the staff scientist who led the study. They did exactly what they were programmed to do, broke when something unexpected happened and couldn’t be changed by anyone who isn’t a robotics engineer.Instead of starting over, the team rewrote those routines in the Model Context Protocol, an open standard that gives an AI a fixed set of specific, pre-written actions rather than open-ended control. With that toolkit, TEM Agent can string steps together, even for requests no one programmed. “It was pretty amazing that it worked out of the box,” Ercius said.Tomography and ptychography on requestResearchers describe an experiment in plain English, and the agent adjusts magnification, tilts the sample, refocuses the beam and chains together dozens of small steps. The team tested it on electron tomography, which rebuilds a 3D view of a sample from tilted images and can take hours. Early on, the agent confused rotating the sample with rotating the scan direction, and a short written explanation fixed it.Stephanie Ribet, a research scientist at the Molecular Foundry, then pushed it into more experimental territory, including ptychography, a technique that reconstructs high-resolution images from diffraction patterns and typically takes significant expertise to set up. Connecting an AI to scientific hardware is harder than connecting it to literature or data pipelines, she said, and the protocol gave the team a way to bridge the two.Linked to supercomputing and a data platformThe agent is also tied to the National Energy Research Scientific Computing Center, a DOE supercomputing facility, and to Crucible, a Molecular Foundry data platform that organizes experiment metadata. Raw data can be processed in seconds instead of hours, and researchers can ask what settings they used on a sample weeks earlier without digging through files or notebooks.Morgan Wall, who leads Crucible’s development, said users can talk in natural language to one interface instead of juggling five scripts on five computers.What comes nextThe approach could extend to other instruments with an open interface, and Wall’s team is helping other nanoscale research centers adopt Crucible. The aim, the release says, is not to replace expertise but to extend it, giving new and experienced researchers a faster path from idea to experiment. Ercius expects such assistants to democratize this type of work. The research was supported by the DOE Office of Science.The work was published in journal npj Computational Materials.