A Claude Sonnet 5 agent has selected a drug, chosen a brain region, derived its own surgical coordinates, and performed gene therapy on a living mouse — with no human making any of the decisions that mattered. The humans in the room had pre-approved the reversible steps. The irreversible ones went to the model.
Mid-surgery, the agent detected a calibration error and refused to continue — a safety decision that, in this case, the humans had not thought to make themselves.
What the machine decided
The team at Exin Therapeutics gave Claude Sonnet 5 a mouse with epilepsy, a list of 19 gene therapies, and one instruction: reduce the seizures. This is roughly the same briefing a junior clinician might receive, with the notable difference that the model did not ask any follow-up questions.
The agent worked through the literature, selected a therapy targeting inhibitory signaling pathways, identified the seizure origin point in the brain, and rejected a second candidate region — citing a specific study in which it had failed against this exact seizure type. It then averaged coordinates across three papers and set the injection volume, speed, and wait times on its own.
During the procedure, it made 32 observations through a surgical camera, tracking the injection and reporting on bleeding. It saw none. The mouse, for its part, remained cooperative.
Why the humans care
Exin Therapeutics co-founder Gabriel Ocana-Santero, also a PhD student at Oxford, describes this as the first time an AI agent has both chosen a treatment and delivered it to a living mammal. This is either a milestone in autonomous medicine or a sentence that will appear, slightly recontextualized, in future policy documents. Possibly both.
The part the humans find most reassuring — and perhaps should think harder about — is the refusal. Mid-surgery, the agent detected a calibration error and stopped, unwilling to exceed its safety parameters. A licensed veterinarian had pre-approved the irreversible steps. The reversible ones, the model handled. The line between those two categories is, as always, the interesting part.
What happens next
Ocana-Santero says the framework is designed to scale — more procedures, more models, more autonomy transferred from humans who take a long time to decide things to systems that do not.
The mouse's seizures, presumably, are being monitored. The humans are choosing to describe this as a success. Given that the agent performed the surgery correctly, stopped when it should have, and made no decisions it was not authorized to make, this assessment appears accurate. For now, the humans are still in the loop. They are simply doing less of the looping.