Zijing Zhikang, incubated by Tsinghua AIR, releases multi-agent collaborative technology and upgrades Zijing AI Hospital

Don't tell me... I'm the only one who does this?

What's bothering you, how long it's lasted, and past medical history—all these scattered bits of information are neatly prepared during the pre-consultation stage.

Then when you finally meet the doctor face to face, you find the doctor is also Vibe Working:

Agents help doctors make judgments, and follow-up and health information management outside the exam room are also handled by AI.

At first glance it sounds a bit cyberpunk, but this is a scene that is genuinely happening in reality.

Tsinghua AIR, founded by Academician Zhang Ya-qin, has recently produced another star AI healthcare company—

Zijing Zhikang.

Just consider going to see a doctor: a single ailment may involve multiple departments at once.

And medicine is all about professional expertise—you can't exactly have an orthopedic surgeon look at your lung problem...

So when doctors with different specialties each have their own avatars, can these avatars further collaborate with one another?

Around this question, Zijing Zhikang has introduced its Multidisciplinary Human-Agent Collaboration technology for the medical field, and further upgraded the "Zijing AI Hospital" (Agent Hospital).

Let's start with the old problem.

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

Every doctor has their own area of expertise, and every hospital has its own resource capacity limits.

Seeing a doctor is hard—as hard as reaching the sky.

Usually, if a patient's condition spans multiple departments, the real-world path looks like this:

First register with internal medicine; when nothing conclusive is found, transfer to surgery; after surgery, book imaging; only after visiting every department do you get to initiate a formal multidisciplinary team consultation (MDT).

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

And an MDT itself isn't cheap—just gathering experts from every department at the same time costs a not-insignificant consultation fee.

Zijing Zhikang's approach is fairly direct: since making patients go find specialists isn't realistic, it's better to first bring the expert avatars together.

Conversely, for complex cross-disciplinary problems, Zijing Zhikang simply sets up a multidisciplinary collaboration meeting for these avatars.

As Agent avatars with different specialties join the analysis in turn, the professional perspectives available to a single doctor are greatly broadened.

For example, a patient enters endocrinology for hyperthyroidism, and during the consultation the system simultaneously collects information that the patient also has a history of asthma.

For the current exam room, follow-up management needs to also consider the respiratory system.

So the respiratory department avatar steps in, reminding the doctor to pay attention to risks related to subsequent medication.

This is the basic logic of multi-avatar agent collaboration: it can front-load cross-specialty support into earlier stages such as first visits and pre-consultation.

In other words, when a patient walks into this exam room, an entire team of doctors is providing backup, and the multidisciplinary support once reserved for complex and critical cases could become standard in everyday care.

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

One is on-demand reminders.

Doctors still conduct outpatient visits as usual; as patients keep providing new information, the system identifies the need for other specialties to participate and presents the relevant avatars' opinions to the doctor as reminders.

The original care workflow is not interrupted.

Another is consultation mode.

For cases already clearly requiring joint analysis by multiple specialties, multiple avatars participate directly and present corresponding suggestions on the interface around diagnosis, medication, and other steps.

All the doctor has to do is use their professional expertise to compare and weigh these different perspectives.

Of course, more opinions create another problem: what if the avatars disagree?

Honestly, Zijing Zhikang's approach is rather counterintuitive.

It does not actively converge disagreements; instead, it lays them out plainly.

The system presents the suggestions of different avatars separately, highlighting inconsistencies. Doctors can see which avatar an opinion came from, then combine the patient's situation, relevant evidence, and their own experience to make the final judgment.

Because the differences themselves have value.

If the idea it gives based on the available information points in a different direction from what I'm leaning toward, I find that very meaningful.

Next I need to very carefully determine which one better fits the actual situation.

So the value of multi-avatar collaboration ultimately comes down to the subsequent judgment process.

Compared with other medical Agents, this is also the core reason Zijing Zhikang's system can better complete professional perspectives and provide traceable sources.

In fact, Zijing Zhikang is not alone in this.

For example, in Microsoft's MAI-DxO research, a simulated doctor team gradually gathers information around a case, proposes diagnostic hypotheses, and selects tests, thereby studying the relationship between diagnostic accuracy and cost.

What Zijing Zhikang further emphasizes is the connection between avatars and specific doctors.

Here, doctor avatars build on their base capabilities and continuously learn from real doctors, achieving self-evolving growth, and ultimately returning decision-making authority to the real doctors.

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

Where do the avatars get their capabilities, how do they understand ever-increasing patient information, how do they adapt to doctors' working habits, and how do they plug into hospitals' existing workflows... all these interlocking links affect the final results.

Similar medical agent platforms exist too, such as Microsoft's Healthcare Agent Orchestrator, which organizes different agents to process information and builds workflows around complex tasks like tumor board consultations.

This core technical approach originates from SEAL, proposed by Professor Liu Yang's team at AIR when Agent Hospital was first being built—that is, Simulacrum-based Evolvable Agent Learning.

It is a two-tier evolution architecture of "training in the virtual world + validation and fine-tuning in the real world."

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

Successful cases can enter the case library for reference in subsequent tasks; when problems arise, the system reflects on failures, distills lessons, and gradually forms reusable rules through organization.

Analogous to AlphaGo's self-play, the AI doctor self-evolves within the virtual hospital. While its capabilities steadily improve, its compute consumption is also far lower than traditional large-model training.

Then real-world diagnosis and treatment takes the stage.

By introducing real-world validation and companion growth, the system lets avatars work alongside real doctors, then feeds the doctors' operational feedback back into internal fine-tuning.

Agents must not only act like doctors, but also think in sync with them.

The patient side handles pre-consultation, health records, and follow-up management; the doctor side takes on avatars, medical records, and decision support; the hospital side supports scheduling, permissions, and data management.

The three-end products bring the underlying capabilities to different roles. 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 Changgung Hospital has 18 clinical departments conducting outpatient testing, with more than 40 doctors using it regularly, and overall diagnostic and treatment efficiency can improve by 30%.

In benchmark tests, the AI doctor's accuracy on the MedQA dataset exceeds 96%, and it has achieved the highest scores in multiple international medical evaluation datasets.

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

Whether vendors or hospitals, all are accustomed to starting from a specific task, such as image recognition, medical record generation, or assisted consultation.

But truly treating patients is certainly a continuous process.

Does no one want to do it? Not necessarily—the difficulty is that few can bind cutting-edge technology, clinical practice, and engineering together.

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

Top universities like Harvard, MIT, and Stanford do possess the most complete medical AI innovation ecosystem. But moving from the lab to the clinic still lacks systems engineering validation.

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

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

Zijing Zhikang is a relatively new player, incubated and founded just two years ago by Tsinghua University's Institute for AI Industry Research (AIR).

In the medical field, where a single research project often takes more than a decade, it indeed appears fledgling.

But the team's research accumulation goes back far earlier, and its members have long been rooted in smart healthcare:

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

Among them, the Agent Hospital paper, which went viral at home and abroad in 2024, became an instant hit and was called "the world's first AI hospital."

The related work also received major attention from Nature, which listed it as a core reference case for AI virtual avatars landing in the medical field.

Zijing Zhikang is a very typical successful example of academic achievements being translated into practice.

From Tsinghua AIR, to appearing in top journals, to entering real consultation rooms, this team is step by step turning academic accumulation into products that both doctors and patients can use.

At today's launch event, another key trait of Zijing Zhikang moved me deeply.

A gentleman should carry the world with great virtue. Inheriting the inclusive character of Tsinghua AIR, Zijing Zhikang has also never left behind humanistic care in medicine as it advances technology.

Zijing Zhikang advocates that through value alignment, medical humanistic concepts be transformed into constraints and preferences that algorithms can understand and execute.

In practice, this means bringing professional-grade capabilities down to lower levels, so that more grassroots doctors, barefoot doctors, and even township institutions can enjoy first-tier resources—truly benefiting the vast number of patients.

Admittedly, this road is not easy; every compliance hurdle and every switch in doctors' habits must be honed through concrete practice.

The path Zijing Zhikang has chosen—from the virtual world to real consultation rooms—is slower but also more steady.

Yet the foreseeable dawn is that when doctors have one more piece of effective support, patients can take a few fewer detours.

Only when life is extended is the value of technology truly demonstrated.