Board people dey always ask me the same question about AI agents: when dem go stop to dey advise and start to dey act? One lab for San Francisco answer that question this week, and the answer come quick pass most governance framework.
Gabriel Ocana-Santero, wey be cofounder and CEO of Exin Therapeutics, one startup wey dey make drug for epilepsy and brain disease, connect one AI agent to surgical machine wey dem build specially. The machine fit hold mouse make e no move, fit watch the procedure through camera, fit move injection needle, and fit control how dem take give the compound. Then im team comot hand.
The agent, wey dey run on Anthropic Claude Sonnet 5, collect spreadsheet wey get nineteen gene therapy options and one instruction only: reduce this mouse seizure. E analyze how seizure dey take happen for that model, pick one therapy wey dey boost one particular signal road for brain, pick the brain part wey science books say na where this kind seizure dey start, and reject the second area with written explanation say im effect no reliable at all. E bring out injection coordinates from three science papers, cross-check the reference points and average the numbers. E set the volume, speed and waiting time by imself. During the procedure, e make thirty-two observations through the camera, track the injection and report whether e see blood. Na veterinary surgeon do the craniotomy; everything after that na the agent own. The mouse dey recover well.
As far as I know, na the first time wey AI agent go choose medical treatment come give am physically to living mammal. Surgical robots don dey more precise for years, but na human being dey always choose the surgery. This time na machine choose.
The thing wey suppose teach person lesson no be the surgery. Na the refusal.
The team first try Anthropic stronger reasoning model, Claude Opus 5.5. E do the experiment planning, then refuse to guide the needle. E reason say to stop for middle with needle inside animal brain wey dem don make sleep bad pass make e no start at all. Sonnet 5, the everyday model, agree to continue. Read am two times. Capability and caution na two different knob. The most capable model for the room no be the one wey gree operate. Any organization wey dey deploy agents inside physical or financial system suppose treat model selection as risk decision, no be procurement decision.
The safety guards too deserve attention, because dem be mechanical, no be mouth. Licensed veterinary doctor approve every step wey no fit reverse. One human hold physical emergency stop button wey fit cut power for all the motors. Dem forbid the agent make e touch the equipment original code. For middle of the procedure, connection cut and spoil the machine coordinate system. The agent notice say the position wey e calculate dey outside the machine physical reach and refuse to continue, three times, even with the animal on top table. The same restriction wey vex the team na im prevent wrong injection. Dem solve am with another agent wey nobody tell say live surgery dey happen.
Na so control plane be. No be policy PDF, but layered restrictions wey go hold when something spoil for the worst time.
Three lessons for every executive wey dey deploy agents pass chat window. First, count where your agents dey touch real world: lab equipment, motor, trading system, industrial control, patient workflow. If you no fit list dem, you no get governance, na hope you get. Second, design for refusal. Agent wey talk no for middle of serious procedure no be fault; na alignment dey work, and your operating model must fit carry that pause without panic. Third, no ever let the operator agent mark im own homework. Exin need second agent wey dem fence from context to repair the first one tools. Separation of duties na old control. E just get new area.
The laboratory of 2030 go pair scientists with agents wey get hands. The companies wey go gain no be the ones wey get the boldest agents. Na the ones wey their agents sabi exactly when to stop, and wey their leaders build stop button before dem build demo.
Think am mathematically. Engineer am well-well. Lead with strategy. Build am responsibly. Scale am globally. Use artificial intelligence create real business value.
