Multi-agent collaboration reshapes medical AI—Tsinghua AIR incubates Doctor A’s digital twin to see patients in hospitals.

Multi-Agent Collaborative Reconstruction of Medical AI! Tsinghua AIR Incubates: Doctor A’s Avatar “Consults” at Hospitals

Now, before each hospital visit, I always let AI run a round for me.

Where it hurts, how long it’s lasted, plus medical history—these bits and pieces are all prepared during the pre‑consultation stage.

When I finally face the doctor, I find the doctor is also Vibe Working:

An agent assists the doctor in making judgments, and AI handles follow‑ups and health information management beyond the exam room.

At first glance it sounds a bit cyberpunk, but it’s a scene actually happening in reality.

Tsinghua AIR, founded by academician Zhang Yaqin, has recently produced another star AI medical company—

Zhijing Zhikang.

Their developed “Zhijing AI Hospital” system has already been rolled out for testing in hospitals across the country.

Even more sci‑fi and futuristic is that every doctor here can have a dedicated AI assistant to support their work, but that’s still far from enough.

Just consider a single visit: one ailment may involve multiple departments simultaneously.

Medicine is also the field that demands the highest professional expertise; you can’t have an orthopedic surgeon diagnose a lung problem…

So when doctors of different specialties each have their own digital avatars, can those avatars collaborate further?

Addressing this question, Zijing Zhikang presents the Multidisciplinary Human‑Agent Collaboration (Multidisciplinary Human-Agent Collaboration) technology in the medical field, and further upgrades the “Zijing AI Hospital” (Agent Hospital).

In AI hospitals, AI doctors collaborate closely with human doctors…a virtual medical world that connects real clinical care and can continuously evolve, is gradually taking shape.

Bringing multi‑department expert judgments into the same examination room

First, let’s address the old problem.

The healthcare industry has an unavoidable reality: good doctors are always scarce.

Each doctor has their own specialty, and each hospital has its own resource capacity limit.

Getting medical care is difficult, as hard as reaching the heavens.

Usually, when a patient’s condition spans multiple departments, the real‑world pathway looks like this:

First you see internal medicine, can’t find the cause so you’re referred to surgery, after surgery you schedule imaging, you’ve booked appointments in several departments, and only then can you initiate a formalmultidisciplinary consultation (MDT).

In the end, the optimal treatment window may already have been missed.

Moreover, a single MDT isn’t cheap; just coordinating the time of experts from each department incurs a substantial consultation fee.

Zijing Zhikang’s approach is straightforward: since it’s unrealistic to have patients seek out experts, why not first bring the experts’ avatars together?

When a condition can be resolved within the same department, a single AI avatar is sufficient.

It’s like inviting an AI replica of a top‑level specialist into a regular clinic to provide guidance and suggestions to the attending physician.

Conversely, for those interdisciplinary complex cases, Zijing Zhikang simply sets up amultidisciplinary collaboration meeting for these replicas.

As agents with different expertise join the analysis one after another, the professional perspective available to a single physician is dramatically broadened.

Thus, under the same operating system’s scheduling,the AI avatar can provide support within the same department and also participate across departments.

For example, a patient admitted to the endocrinology department for hyperthyroidism also has a history of asthma, which the system captures during the consultation.

For the current examination room, subsequent management must also take the respiratory system into account.

Thus, a respiratory specialist avatar intervenes, reminding doctors of the related risks of subsequent medication.

This is the basic logic of multi‑avatar intelligent agent collaboration, which canbring cross‑specialty support forwardto earlier stages such as the initial consultation and pre‑triage.

In other words, when a patient walks into this examination room, an entire medical team is already providing assistance, and the specialized multidisciplinary support for complex cases may become the norm in everyday clinical practice.

How exactly does the collaboration work? Zijing Zhikang provides two forms.

One is on‑demand reminders.

The doctor continues with regular outpatient visits, and as patients provide new information, the system identifies the need for involvement from other specialties and then presents the relevant consults as reminders to the doctor.

This does not interrupt the existing diagnostic and treatment workflow.

Another approach is the consultation mode.

For cases that clearly require analysis by multiple specialties, multiple avatars directly participate, and the interface presents corresponding recommendations around diagnosis, medication, and other steps.

What the doctor needs to do is use their professional knowledge to weigh these different perspectives.

Of course, with many opinions, another problem arises:the different personas say different things, what to do?

Honestly, Zijing Zhikang’s approach is quite counterintuitive.

It doesn’t actively try to converge the divergences; instead, it lays the differences out clearly.

The system presents the recommendations from each persona separately, highlighting any inconsistencies. The doctor can see which persona each opinion comes from and, together with the patient’s situation, relevant evidence, and personal experience, make the final judgment.

Because the differences themselves are valuable.

A practicing internist said that, at present, AI’s important assistance is still to supplement directions that doctors may not have considered:

When AI’s suggestions don’t fully align with my judgment, I review the input information to see what caused the discrepancy.

If the ideas it proposes based on existing data are in a different direction from what I tend to favor,I find that very meaningful.

Then I need to carefully determine which one actually reflects the reality.

Therefore, the value of multiple avatars collaborating ultimately depends on thesubsequent assessmentprocess.

Compared with other medical agents, this is also the core reason why Zijing Zhikang’s system can better supplement professional perspectives and provide traceable sources.

In fact, this is not only Zijing Zhikang’s claim.

Organizing multiple AI personas has already become one of the key research directions in medical AI.

For example, in Microsoft’s MAI‑DxO study, they simulate a team of doctors who gradually gather information about a case, propose diagnostic hypotheses, and select tests, to investigate the relationship between diagnostic accuracy and cost.

ZiJing ZhiKang further emphasizes the connection between the avatars and specific physicians.

The doctor avatars here will build on their basic capabilities, continuously learn from real doctors, achieve self‑evolutionary growth, and ultimately return decision‑making authority to the real doctors.

From this perspective, the collaboration of multiple avatars actually surpasses the traditional narrative of “using AI to replace doctors”; instead, it means that behind every doctor stands an entire on‑demand expert team.

Foundation Formed, Enabling AI Avatars to Continuously Grow

However, to keep hundreds or thousands of avatars online at all times, each performing its own duties, a complete system is required behind the scenes.

Where do the avatars get their capabilities, how do they interpret the ever‑growing patient information, how do they adapt to doctors’ work habits, and how are they integrated into existing hospital workflows… each link is interconnected and will affect the final user experience.

This is also why Zijing Zhikang is building the Zijing AI Hospital Ecosystem.

There are similar medical agent platforms, such as Microsoft’s Healthcare Agent Orchestrator, which coordinates different agents to process information and builds workflows around complex tasks like tumor consultations.

What sets Zijing AI Hospital apart is that its avatars continuously evolve through real outpatient practice.

This core technical roadmap originates from the SEAL proposed by Professor Liu Yang’s AIR team at the inception of building Agent Hospital, which stands for evolvable agent learning based on simulacra.

It is a dual‑layer evolutionary architecture of “virtual world training + real‑world validation and fine‑tuning”.

In the virtual environment, roles such as patients, doctors, and nurses are all played byintelligent agents.

AI doctors conduct high‑intensity simulated diagnosis and treatment daily, with millions of ultra‑realistic AI patients seeking medical advice, forming a closed‑loop virtual medical world.

Successful cases can be added to the case library for future reference; When problems arise, they are reflected upon and distilled from failures, and after organization, gradually form reusable rules.

In just a few short days, the AI doctor agent can complete hundreds of thousands, even millions of cases, running through all rare diseases, complex diseases, and multiple disease-course combinations.

Analogous to AlphaGo’s self‑play, the AI doctor evolves autonomously in a virtual hospital. While its capabilities steadily improve, the computational cost is far lower than traditional large‑model training.

However, early AI trained in the virtual world has obvious issues; it tends to consider every possibility overly exhaustively.

Qiao Yuchen, Chief Product Officer of Bauhinia AI Hospital, revealed this interesting detail:

The AI recommends the most effective medication to doctors, but the doctor, considering the patient’s financial situation, ultimately switches to a relatively cheaper, insurance‑covered drug.

Thus real‑world treatment steps in.

By introducing real-world validation and co‑evolution, the system lets the avatar work alongside real doctors, and feeds the doctors’ actions back into internal fine‑tuning.

The agent not only acts like a doctor, but its thinking must also be synchronized with doctors.

In practical application, the Zijing AI Hospital serves three sides: doctors, patients, and hospitals:

Patient side handles pre‑consultation, health records, and follow‑up management; Doctor side takes on avatars, medical records, and decision support; Hospital side supports scheduling, permissions, and data management.

The three‑side products bring underlying capabilities to each role. Continuing to integrate specialty tools and partners is the next step built on this foundation.

Currently, this system has already delivered results in real hospitals.

As of July 2026, the system has built 54 AI doctors and millions of hyper‑realistic AI patients, covering thousands of diseases.

Beijing Tsinghua Chang Gung Hospital already has 18 clinical departments conducting outpatient tests, with over 40 doctors using it routinely, and overall physician diagnostic efficiency can be improved by 30%.

In benchmark testing, AI Doctor achieves an accuracy exceeding 96%, and attains top scores across multiple international medical evaluation datasets.

The next question is how to bring this professional support into more hospitals and benefit more patients.

The next step for AI hospitals: reaching more people

In recent years, medical AI has tended to operate in isolation.

Whether it’s vendors or hospitals, they’re accustomed to starting with a specific task, such as image recognition, medical record generation, or assisted consultation.

But when it comes to actual diagnosis, it’s definitely a series of processes.

The industry has long lacked a production‑grade AI‑native system that can connect different roles and span the entire clinical workflow.

Does that mean no one wants to build it? Not necessarily; the difficulty lies in the fact that few can tie cutting‑edge technology, clinical practice, and engineering together.

This is the awkward situation faced jointly by academia and industry:

Harvard, MIT, Stanford and other top universities, indeed have the most comprehensive medical AI innovation ecosystem. However, moving from experiments to clinical practice still lacks systematic engineering validation.

Top medical institutions(such as Mayo Clinic, Toronto General Hospital, Cleveland Clinic) do have deep clinical resources and abundant patient demand.

But they lack top AI talent and technology, so they cannot independently develop AI-native systems.

AI tech giants in Silicon Valley may go a bit farther, but medicine is just one of many areas they are venturing into; they lack system-level innovation and need to gradually understand and adapt to medical workflows in real clinical practice.

The deployment of production-grade AI-native medical systems requires these capabilities to work together, so currently Zijing Zhikang is at T0 level.

Zijing Zhikang is a new player, having been incubated and founded byTsinghua Intelligent Industry Research Institute(AIR) just two years ago.

In a medical field where single‑topic research often spans decades, it indeed looks like a newcomer.

But the team’s research foundation predates this by far, and its members have long been rooted in smart healthcare:

Multiple papers have consecutively appeared in top medical AI journals such as Nature Medicine and NEJM AI.

Among them, the Agent Hospital paper that went viral worldwide in 2024 shot to fame and was dubbed the “world’s first AI hospital”.

The related work also received prominent attention from Nature and was listed as a core reference case for AI virtual avatars in the medical field.

Zijing Zhikang is a very typicalsuccessful example of academic成果落地.

Starting from Tsinghua AIR, reaching top journals, and entering real clinics, this team is gradually turning academic accumulation into products that both doctors and patients can use.

One More Thing

At today’s launch event, Zijing Zhikang displayed another key quality that moved me deeply.

A gentleman should bear the world with profound virtue, inheriting the inclusive qualities of Tsinghua AIR, Zijing Zhikang has never neglected the humanistic care of medicine throughout its technological advancement.

Zijing Zhikang advocates that, through value alignment, medical humanities concepts can be transformed into algorithmically understandable and actionable constraints and preferences.

In other words, the system is designed from the outset to firmly define the collaborative boundary between AI and humans, adhering to a human‑centric approach and clearly positioning AI as an aid.

In building agents, we alsodo not pursue an all-powerful, unified super AI, but retain the differences and diversity of each agent, drawing on the strengths of many.

In practice, this means delegating professional‑level capabilities so that more primary care doctors, barefoot doctors, and even township institutions can access frontline resources, truly benefiting a broad patient base.

Admittedly, this path is not easy; every compliance hurdle and every shift in doctors’ habits must be refined through concrete practice.

The path that Zijing Zhikang has chosen from the virtual world to the real clinic is slower but steadier.

However, the foreseeable dawn is that for each additional effective support a doctor receives, patients can avoid some detours.

When life can be prolonged, the value of technology is truly demonstrated.

Official website link: https://www.tairex.cn/

One-click Triple Action“Like” “Share” “Heart”

Feel free to leave your thoughts in the comments!

— End —

🌟 Light Up the Star ⭐️

Daily Updates on Cutting‑Edge Tech