Tsinghua AIR Spinout Upgrades Its AI Hospital With Multi-Persona Agent Collaboration

Zijing Zhikang, a company incubated by Tsinghua University's Institute for AI Industry Research, has upgraded its Agent Hospital system with what it calls Multidisciplinary Human-Agent Collaboration, according to a QbitAI report published on October 21.

The system is already being tested in hospitals across China, the report says. Each doctor gets a dedicated AI persona: a single persona is enough for problems that stay within one department, while complex cross-disciplinary cases bring several specialist personas into the same consultation.

There are two collaboration modes. In an on-demand mode, the doctor runs the clinic as usual while the system watches for signs that another specialty should weigh in, surfacing that persona's opinion as a prompt without interrupting the workflow. In consultation mode, multiple personas participate directly, offering suggestions on diagnosis and medication.

One design choice runs against intuition: Zijing Zhikang does not try to collapse disagreement between personas. Their recommendations are presented separately, inconsistencies are highlighted, and the doctor can see which persona produced which view before making the final call. The report quotes an internal medicine physician saying that moments when the AI disagrees are exactly when it is worth revisiting the inputs and discriminating carefully.

The underlying method is SEAL, the evolvable-agent learning approach the AIR team proposed when it first built Agent Hospital. Inside the virtual hospital, patients, doctors and nurses are all played by agents; AI doctors run high-volume simulated consultations, successful cases enter a case library, failures are reflected on and distilled into reusable rules, and real clinics then supply validation and feedback for fine-tuning.

Numbers cited in the report run through July 2026: 54 AI doctors and millions of simulated patients covering thousands of diseases. At Beijing Tsinghua Changgung Hospital, 18 clinical departments have run outpatient tests with more than 40 doctors using the system routinely, and overall consultation efficiency is described as 30% higher. The report also says the AI doctors exceed 96% accuracy on the MedQA dataset and top several international medical evaluation sets.

The product is split across three surfaces: a patient side for pre-consultation, health records and follow-up; a doctor side for personas, medical records and decision support; and a hospital side for scheduling, permissions and data management. Connecting specialty tools and partners is described as the next step.

The signal here is about system boundaries. Medical AI has mostly arrived as single-task tools for imaging, note generation or triage, whereas this system pushes cross-specialty support into the first visit and pre-consultation stage while leaving the final decision with the human doctor. What to watch is whether the hospital footprint keeps expanding, how disagreement and liability get defined in practice, and whether capabilities learned in simulation hold up in real clinics.