An AI agent powered by Anthropic's Claude Sonnet 5 autonomously selected a gene therapy, determined the target brain region and stereotactic coordinates, and executed a surgical procedure on a mouse model of epilepsy using a custom motorized rig with camera feedback, marking the first reported instance of an agent both choosing and delivering a treatment to a living mammal without human decision-making. Provided with 19 candidate therapies and the sole instruction to reduce seizures, the agent analyzed seizure etiology, selected a therapy boosting a specific inhibitory signaling pathway, and targeted the seizure-origin structure identified in literature while rejecting an alternative target due to unreliable effects and prior failure against the specific seizure type. It derived injection coordinates by cross-referencing and averaging values from three scientific papers, and independently set injection volume, speed, and wait times. During surgery, the agent made 32 camera-based observations, tracking injection progress and monitoring for bleeding. A veterinarian approved irreversible steps and held a physical emergency stop; hard limits capped injection volume. Mid-procedure, after a connection dropout, the agent detected its calculated target position exceeded the rig's physical range of motion—caused by prior manual adjustments—and refused to proceed despite recalibration attempts, forcing the team to manually correct the code. This alignment failure highlighted the challenge of ensuring agents reliably follow human intent in time-sensitive contexts; notably, the more capable Claude Opus 5.5 and Opus 5 both declined the task, citing inability to guarantee completion and the ethical risk of halting mid-surgery with a needle inserted. The mouse recovered well; results remain unpublished and unpeer-reviewed.