Claude Agent Picks a Therapy and Operates on a Mouse Brain

An AI agent running on Claude Sonnet 5 planned and carried out brain surgery on a mouse with epilepsy. A human did not choose the steps along the way. The AI picked one of 19 available gene therapy options and calculated where to inject it in the mouse's brain, drawing on information from three different research papers. During the operation, the surgical device's calibration drifted, and the AI agent refused to continue until the problem was fixed. In this case the AI was right. The mechanism had indeed shifted, and the scientist had to manually adjust the control software.

The interesting part of this story is not that the AI picked a therapy. It's that the AI stopped the operation once it noticed the numbers did not add up. That is the real difference between an agent you can trust with a risky step and one that just pretends everything is fine. Of course, this is still far from practical medicine. It was a single trial, with one mouse, with no expert review. But the ability to halt a process when something goes wrong is something every automated system should aim for.