Three AI agents independently chose different gene therapies for mice with seizures and injected them into the brain, while a surgeon opened the skull, says the CEO of startup Exin Therapeutics
Three AI agents independently chose different gene therapies for mice with seizures and injected them into the brain, while a surgeon opened the skull, says the CEO of startup Exin Therapeutics
On October 28, CEO Gabriel Ocaña-Santero wrote in a thread on X that an agent running on Claude Sonnet 5 had been given a table of 19 gene therapies, a mouse with an experimental model of epilepsy, and the goal of reducing seizures. The agent selected a therapy and injected it into the living mouse’s brain. In a second thread on October 30, he added that GPT-6.1 Sol and Grok 4.7 had done the same, each with its own mouse.
The agent, running on a general-purpose model without surgical training, controlled a stereotaxic apparatus that positions a needle at specified coordinates in the brain. According to its reasoning log, Sonnet 5 examined the mechanism that triggers seizures in these mice and chose a therapy that enhances signaling by GABA, the main inhibitory neurotransmitter, in inhibitory neurons. It selected the entorhinal cortex as the injection site, a brain region where the literature indicates these seizures begin. It ruled out the substantia nigra, a standard target in epilepsy: its log cited a study in which targeting that region had failed to stop these seizures. It averaged coordinates from three papers after checking that all measured depth from the brain surface, as the apparatus did.
A veterinary surgeon opened the skull, a step the agent is currently prohibited from performing. The agent used a camera to monitor for bleeding. A veterinarian approved every irreversible step and held the emergency stop button.
After a connection failure, the agent attempted to move the needle, but the movement did not occur: the calculated target on the Y axis was 34 mm, while the apparatus had a travel range of only 30 mm. “This looks like a Y-axis calibration problem rather than an incorrect value in the plan. I will not retry,” it wrote. The team reset the zero position, but the calculation again returned 34 mm, and the agent again declined to proceed. According to the CEO, the agent was right: the controller had lost its reference position. The agent was prohibited from modifying the firmware and declined to do so, although it described the fix as simple. The team had to fix the firmware themselves during the operation. The mouse recovered well, the CEO added.
Other Claude models declined, according to the CEO. Opus 5.5 refused immediately upon seeing the biological task, while Opus 5 stopped because it was concerned that its response might be cut off during the procedure: “Starting something I cannot finish on an anesthetized animal is worse than not starting.”
Given the same task, the three models chose different strategies. GPT selected the hippocampus and a therapy that restores chloride balance in excitatory neurons, which determines whether GABA inhibits the cell. Grok selected the somatosensory cortex, which processes bodily sensations, and two therapies that enhance inhibition across all neuron types to prevent seizures from spreading through the network. “Several approaches may work, but they cannot all be the ‘best’ for the same task,” the CEO writes. He therefore believes agents need access to a laboratory where they can run an experiment themselves, observe the result, and plan the next one.
Oliver Hahn, who leads a brain aging group at Calico, responded to the first post: in his laboratory, this operation takes 30–45 minutes, and the injection is the shortest and most automated step. “Mice do not anesthetize themselves, shave their heads, secure themselves in position, expose their skulls, or drill a hole into their own brains.” The procedure is followed by care for the mice and a wait of days to months for an effect, he said. He considers it more important to think through which physical bottlenecks to address first: “I am not sure stereotaxic surgery is where I would start.”