A biotech founder equipped an artificial-intelligence agent with a robotic arm, injector and a custom-built surgical rig to perform a brain injection on a living mouse. The setup included a motorized device that immobilized the animal, a camera that gave the agent visual feedback, and a needle that could deliver a compound. The experiment tested a gene-therapy candidate intended to reduce seizures in a mouse model of epilepsy.
Santero said the goal was to let AI agents interact directly with laboratory hardware rather than limiting them to experimental design or data analysis. He explained that models trained on published literature often miss the nuances of real-world research, such as failed attempts, unexpected observations and on-the-fly decisions. Integrating AI with actual experiments, he argued, could make the systems more reliable and useful for drug discovery.
The team connected the AI to a rig that could observe a veterinary surgeon performing a craniotomy, then let the model choose a therapy, brain region and injection coordinates from a spreadsheet of nineteen options. After the surgeon opened the skull, the agent directed the robotic arms, set the injection volume and speed, and guided the needle into the target area. The first model tested, Anthropic’s Claude Opus 5.5, declined to advance the needle, citing a risk of stopping mid-procedure.
Human safeguards included a requirement that the AI submit an injection plan for human review and an emergency button to cut motor power. The agent’s access to control code was limited to prevent self-modification. When a software glitch shifted the coordinate system, the AI flagged the error but would not change the code because its instructions forbade it. A separate Opus 5 instance patched the code, unaware that a live surgery was occurring.
The combined procedure lasted just over an hour, longer than the typical forty-five minute brain injection. The team is now monitoring the mouse to see whether the gene-therapy reduces seizure activity. Despite the delays, the researchers said the AI performed the task competently and plan to evaluate other models such as Grok. Future tests will also explore agent-controlled craniotomies, a more complex operation that the team wants to perfect before any further live experiments.
Santero believes AI agents will increasingly act as collaborators alongside human scientists, with direct control of laboratory instruments. He stresses that human oversight remains essential, especially when experiments involve living animals, and cautions that deploying such systems on human brains is far premature. The demonstration highlights both the potential and the current technical and ethical challenges of embedding artificial intelligence in hands-on biomedical research.