After several weeks of hard work training agents, the ACEM Core realized on October 22, 2026 a loop in which two AI agents work together and repeat until a task is done. Both agents run on the core’s microscope computers. One agent plans the experiment and the other runs it on the microscope. They report back to each other after every run.
The first task was to find the focus on a sample. The instrument agent ran a focus recipe on the Spectra Ultra: it opened the column valves, started the FluCam view, stepped the beam through three magnifications and swept the stage height. The recipe found focus at z = 53.9 µm, down from a starting height of 68.9 µm, and finished with the beam blanked. The planner then checked the result and the 4800× calibration, confirmed that all the safety steps had worked, and began drafting the next plan, a TEM survey of the sample.
The loop also handles instructions from the person at the microscope. When the user gave a new rule, “always close the column valve while waiting”, the planner passed it to the other agent. The instrument agent added it to the recipe so that the valves are closed at the end of every run and during every pause, and reopened only after the standard checks.
This opens the door for AI to evolve itself: the agents improve their own recipes from what they learn in each run. Human input is still required for more advanced tasks. The core is building on this work as part of the AI and automation program, with the aim of giving users agents that can plan, run and check experiments safely, with a person in charge throughout.