Exin Therapeutics conducted experiments on live mice with epilepsy. An AI agent screened 19 gene therapy candidates, determined brain targets and injection coordinates, and controlled the equipment to perform the brain injections. A veterinary surgeon carried out the craniotomy for the experiment, with manual safety safeguards in place: the veterinarian could approve irreversible steps and trigger an emergency stop, and the AI was prohibited from modifying the device's low-level control code. The experiment tested two large models in succession; during the process, a coordinate system failure occurred and was resolved by another independent AI agent that modified the code. The overall process took longer than conventional injection surgery. Currently, results related to epileptic seizures have not been published, nor have they undergone peer review. This experiment is a human-supervised demonstration of AI-led injection and represents an important step for AI—from data analysis toward interfacing with physical research tools. Such AI agents connected to physical experimental equipment can form feedback loops, potentially accelerating the efficiency of hypothesis testing, but they also place higher demands on reliability and human oversight.