What a clinical agent should see

The first generation of medical AI was often presented as a question-answering system. The next generation has to be closer to the work.

Doctors do not reason in a blank input field. They move across images, prior notes, orders, documents, lab trends, messages, and patient histories. They compare what they are seeing now with what happened before. They ask whether the missing information changes the next step.

The workspace is part of the model

In SaroviX, the agent should know which patient is open, which imaging study is visible, which slice or region is being discussed, what the prior note said, and what the doctor is trying to produce. That context makes the agent less theatrical and more useful.

A radiology question should be grounded in the DICOM/NIfTI view. A note-writing task should be grounded in the encounter and prior timeline. A molecular question should be grounded in the patient baseline and protein or pathway context. A care-team question should be grounded in tasks, red flags, follow-up, and who is responsible.

Transcript outline

First, the agent listens. It should understand spoken intent without forcing the physician into unnatural commands. Second, the agent looks. It should read the active workspace state: scan, document, note, task, patient, and relevant history. Third, the agent acts with permission: draft the SOAP note, compare a scan, produce a patient summary, or ask for missing data.

Fourth, the agent leaves a trail. Medical AI cannot be a magic sentence machine. It should show what it used, what it inferred, and where uncertainty remains.