What they did
Ocana-Santero, cofounder and CEO of Exin Therapeutics and a PhD student at Oxford, set up a mouse model of epilepsy and a custom motorised stereotactic device with a microscope, cameras and a syringe. Claude Sonnet 5 was given a list of 19 candidate gene therapies, control of the device and one instruction: reduce the seizures. It worked out how seizures start in this model, chose a therapy that boosts an inhibitory pathway, picked the brain structure the literature names as the seizure origin and argued against a second region. It took injection coordinates from three papers and averaged them, and set volume, speed and wait times. A licensed vet approved every irreversible step and held a hardware emergency stop, and fixed limits such as an injection volume cap were enforced.
How it went
The agent carried out the injection, logging 32 observations from the camera feed, including progress and whether it saw bleeding. The mouse recovered well. After a glitch it found the computed target position was wrong and refused to go on, because its instructions barred it from editing the rig's code. The two accounts differ on the fix: Decoder says Ocana-Santero fixed the code by hand with the animal under anaesthesia, while WIRED reports a second agent on Opus 5, not told a live surgery was underway, made the change. The procedure ran a little over an hour against about 45 minutes for a typical one. Whether the seizures dropped is not reported yet.
Worth knowing
Two outlets reported this, both from Ocana-Santero's own account; no technical report or peer review is out yet, and the harness was not named. Sonnet 5 was the only model that would run the live procedure; Opus 5.5 and Opus 5 refused, saying stopping midway with a needle in the brain would be worse than not starting. Plan for the agent halting at the worst moment. The team plans to test Grok and other models next.