On October 21, Zijing Zhikang, a two-year-old startup incubated by Tsinghua's Institute for AI Industry Research (AIR), held its "Zijing AI Hospital Multi-Agent Collaboration Ecosystem Launch" at the Beijing National Convention Center. It released a medical multi-avatar agent collaboration technology (Multidisciplinary Human-Agent Collaboration) and upgraded the Zijing AI Hospital system — answering a question few had tackled: when every doctor has an AI avatar, can those avatars work together?

Futuristic clinic illustration with neural networks and medical icons
Multi-avatar collaboration brings a specialist team to the first visit

Bringing MDT to the First Consultation

The real-world path for complex cases is long: internal medicine first, then surgery, then imaging — and only later a formal multidisciplinary team (MDT) review, by which time the best treatment window may have closed. Zijing's answer is to assemble specialist avatars upfront: ① same-department support, where an expert avatar joins ordinary clinics to advise the attending doctor; ② cross-department collaboration — when a hyperthyroidism patient also has a history of asthma, a respiratory avatar steps in to flag medication risks.

Disagreements Surfaced, Not Suppressed

When avatars disagree, the system does not force a consensus. It presents each recommendation side by side, highlights inconsistencies, labels the source of every opinion, and hands the final call back to the physician. One internist put it plainly: AI's value is revealing directions a doctor may have missed — "when AI points the other way, that is exactly when I must discriminate most carefully."

Trained in a Virtual World, Evolving in the Real One

The foundation is the SEAL architecture (Simulacrum-based Evolvable Agent Learning) proposed by Professor Liu Yang's team at Tsinghua AIR: inside a virtual hospital, AI doctors run hundreds of thousands to millions of simulated cases within days, then keep evolving through validation and fine-tuning in real clinics — learning, for instance, to weigh the "best-efficacy" drug against what medical insurance actually covers. As of July 2026 the system had 54 AI doctors and millions of simulated patients spanning thousands of diseases; at Beijing Tsinghua Changgung Hospital, 18 clinical departments are piloting it, 40+ doctors use it routinely, and overall diagnostic efficiency is up 30%. AI doctors exceed 96% accuracy on MedQA, and the online virtual clinic has been open for global public beta since April 2026.

Zijing AI Hospital by the numbers (as of July 2026)

· 54 AI doctors covering thousands of diseases; millions of simulated patients

· Tsinghua Changgung Hospital: 18 departments, 40+ doctors in routine use

· Diagnostic efficiency up 30%; MedQA accuracy >96%

Medical AI is shifting from single-point tools to workflow-level platforms — Microsoft's MAI-DxO research and Healthcare Agent Orchestrator point the same way. What distinguishes the Zijing path is avatars growing alongside real physicians, and a deliberate push to bring front-line expertise to primary care. Multi-agent collaboration is also the design doctrine of NineZenith's Zenith-Control platform: AI offers perspectives; humans make decisions.

(Information synthesized from QbitAI, Pedaily and Xinhua reports)