An AI agent picked the treatment and target, then performed mouse brain surgery

A Claude Sonnet 5 agent chose a gene therapy and brain target for epileptic mice and operated, under a vet's supervision. It is unreviewed and self-reported.

An AI agent running on Claude Sonnet 5 chose a gene therapy, a brain region and injection coordinates for a mouse epilepsy model, then carried out the surgery. Gabriel Ocana-Santero, an Oxford PhD student and cofounder and CEO of Exin Therapeutics, described the work in a thread on X. The agent got 19 candidate gene therapies, a mouse model with induced seizures and a custom motorized surgical device with cameras. Its only instruction was to reduce the seizures.

The agent proposed a therapy that boosts an inhibitory signaling pathway. It picked a brain structure the literature names as the origin of this seizure type, and it rejected a second region because of unreliable effects and a study where that region failed against the same seizure type. It derived coordinates from three papers, cross-checked reference points and averaged them, and set injection volume, speed and wait times itself. During surgery it made 32 camera-based observations, such as reporting bleeding and injection progress.

Humans stayed in the loop. A licensed veterinarian approved each irreversible step and held a physical emergency stop that cuts power to all motors, and fixed limits like an injection volume cap could not be overridden. After a dropped connection, the agent judged its target to be outside the device's physical range and refused to continue, even after a recalibration and a third attempt. It was right: manual adjustments during an earlier pause had thrown off the control system, and Ocana-Santero fixed the code by hand. He says the mouse recovered well.

Why it matters: gene-therapy research leans on scarce specialist labor for choosing candidates and for precise surgery. If agents can do both, even partly, testing ideas in animals could get faster and cheaper, and that is a precondition for cheaper treatments later. This is a single mouse procedure, not a workflow.

Caveats: the results are not peer-reviewed, and a technical report is expected but unpublished. The first-of-its-kind claim is Ocana-Santero's own, and the account comes from his X thread without independent verification. He also says Claude Sonnet 5 was the only model willing to run the live procedure; Opus 5.5 and Opus 5 refused.

Source: The Decoder