Gabriel Ocana-Santero, who runs Exin Therapeutics and studies at Oxford, said on X that a Claude Sonnet 5 agent treated an epileptic mouse. The Decoder reported his account on 30 October. The team gave the agent 19 gene therapies to choose from, a seizure model and a custom motorized surgical rig with cameras. Its one goal was fewer seizures. He says the agent picked the therapy, brain region and coordinates itself. A licensed vet approved each step that could not be undone and held a power cutoff.
After a dropped connection, the agent judged that its target lay outside the rig's range of motion. It refused to go on, even after recalibration and a third try. Ocana-Santero says the agent was right, because hand adjustments during an earlier pause had thrown the control system off. He fixed the code by hand. He says the mouse recovered well. No peer review has taken place, and a technical report is still to come. Opus 5 and Opus 5.5 had refused the job. Opus 5 said stopping with a needle in an anesthetized brain was worse than not starting.
That is the tension: the one model willing to start was the one that stopped midway. Teams using agents should ask who can override a refusal, how state is rechecked after hand changes, and what happens to a halted job.