The Evolvable AI Doctor Is Here! How Tsinghua-Spawned Agent Hospital Managed to Penetrate Hospitals Across China

As the patient sits down and describes their condition, the system operates seamlessly in the background: converting speech to text, extracting information, and generating a draft medical record. By the end of the consultation, a well-structured electronic medical record is already displayed on the screen, awaiting the physician’s verification, confirmation, and signature.

At first, the doctor might have merely found it convenient. A few weeks later, the diagnostic suggestions provided by the AI avatar increasingly aligned with his own clinical reasoning, even proactively reminding him when he overlooked a key segment of the patient’s medical history.

This digital avatar originates from Tairex’s AI Hospital system (i.e., the Agent Hospital system)—its technological core is precisely the Agent Hospital proposed by the Tsinghua University team in their 2024 paper.

In May 2024, the team from the Institute for AI Industry Research (AIR) at Tsinghua University published a paper titled “Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents,” proposing Agent Hospital, a virtual medical world composed of evolvable intelligent agents. In September 2024, Tairex was established as a spin-off from the Institute for AI Industry Research at Tsinghua University.

Nearly two years later, in July 2026, Nature published the article “AI avatars are reshaping society in China- the law is trying to catch up” online, citing Agent Hospital as a core case study of the clinical implementation of AI virtual avatars in China’s healthcare sector and listing the aforementioned paper as its sole reference.

From the experimental concepts outlined in academic papers to practical application in consultation rooms across numerous hospitals in China, the Tairex AI Hospital System has evolved from an academic testing ground to real-world clinical settings. Based on industry practice outcomes, the future development of the Tairex AI Hospital System will not merely involve adding a few AI tools to traditional hospitals, but rather making high-quality medical capabilities replicable and scalable.

Typically, medicine is regarded as an empirical discipline. It takes over a decade for human physicians to progress from medical students to attending physicians. Experience accumulates over time, and case exposure requires prolonged waiting. Rare diseases may be encountered only a few times in a lifetime.

The Tairex AI Hospital System can accelerate time by 100 times by establishing a virtual medical world.

So, how can the problem of data scarcity in AI healthcare be addressed?

Real medical records involve privacy concerns and cannot be discharged on a large scale; even within the same hospital, the vast amount of unstructured information in electronic medical records leaves algorithms at a loss. Data from the National Health Commission shows that the usage rate of electronic medical records in tertiary hospitals across China is 98%, but only 28% of the content is in a standardized format that computers can directly interpret.

In the virtual clinic, AI doctors interact with over one million AI patient agents, covering 26 clinical specialties and thousands of diseases, with diverse variations in age, gender, geographic location, and medical history. Leveraging foundation models integrated with authoritative medical knowledge bases, these agents are automatically generated through a designed hierarchical architecture and quality assessment framework. This approach offers a novel solution to the longstanding scarcity of medical data, previously constrained by privacy protection concerns.

The aforementioned paper data shows that after diagnosing and treating approximately 10,000 patients in a virtual environment, the AI doctor achieved an accuracy rate of 93.06% on the MedQA respiratory disease subset, whereas it would take human doctors more than two years to complete the same patient volume.

In the temporal dimension, AI physicians can complete diagnostic and therapeutic practice equivalent to years or even decades of human clinical experience within a matter of days. Professor Liu Yang’s team refers to this phenomenon as the “Scaling Law of Evolution”: the more AI patients an AI physician treats, the greater its capabilities become. If AI patients meet the criteria of being hyper-realistic, broadly distributed, and diverse, the capability evolution curve adheres to this scaling law. The AI physician developed by his team achieved an accuracy rate exceeding 96% on the internationally authoritative MedQA dataset (the benchmark for the United States Medical Licensing Examination) by the end of 2024.

AI patients generated by the system exhibit strong anthropomorphic capabilities, possessing distinct professional identities and emotional traits, and can even proactively send test orders to physicians.

Thus, AI doctors can self-evolve in the virtual world.

AI doctors must not rest on high scores achieved in laboratory settings; the true test begins when they step out of the virtual world and into real clinical practice.

There is no shortage of medical AI products on the market. Image recognition, medical record transcription, consultation assistants... However, general-purpose large language models deployed in hospitals have received feedback indicating that their responses are rather generic and lack depth.

One of the advantages of Tairex's digital twin agents is their "evolvability."

In addition to completing "high-throughput evolution" in the virtual world, symbiotic evolution is achieved through linkage and interaction with the real world.

Traditional medical AI is often statically deployed, unable to iterate once it leaves the development environment. In contrast, Tairex’s Physician Avatar Agents are mature models pre-trained on high-throughput data from specific departments during initialization. Once deployed in real-world hospitals, they silently observe and record every human physician’s consultations, prescription modifications, and diagnostic decisions. The system continuously performs adaptive fine-tuning based on minor discrepancies between the AI’s recommendations and the physicians’ final decisions, a process that requires no additional cumbersome data annotation or manual intervention from doctors.

At this stage, the AI physician resembles an apprentice accompanying a supervising mentor on clinical rounds, quietly growing with each consultation.

A frontline physician shared their feedback after practical testing: “Some AI-generated recommendations were immediately recognizable as ‘fluff’ to professionals. But now, they have reached a level where I find them usable.”

A pulmonologist once shared a case: an elderly patient who had undergone coronary stent implantation presented with chest tightness. During the initial consultation, the physician focused primarily on ruling out respiratory causes, while the AI assistant suggested that the chest tightness might be related to the stent itself rather than a respiratory issue. In clinical practice, specialists who have long been deeply engaged in a specific field may sometimes overlook potential influences from other body systems. In such scenarios, AI-assisted systems provide valuable supplementary reference.

On April 10, 2026, Tairex officially launched the "Tairex AI Hospital," an online virtual clinic, opening applications for testing to licensed physicians and medical students worldwide. The system features 26 clinical departments in Chinese and 21 in English, each populated with large-scale, highly anthropomorphic, and broadly distributed AI patient agents. To date, domestic users span 27 provinces, municipalities, and autonomous regions across China, while international users come from multiple countries and regions, including Italy, Russia, the United Kingdom, Turkey, and Ireland, comprehensively covering all departments offered in the online virtual clinic. Within this virtual environment, after completing a consultation, users can view the distribution of diagnostic and treatment decisions made by other physicians for the same AI patient, facilitating mutual learning and reference.

Avatar agents rapidly accumulate experience by diagnosing and treating a massive volume of AI patients over short periods. Their initialization baseline is already that of a mature, highly capable agent pre-trained on department-specific data. Subsequently, in real-world clinical settings, the avatar agent collaborates with human physicians when they utilize the system. Notably, at Tairex AI Hospital, AI systems are primarily responsible for information collection, analysis, and generating decision-support recommendations, while all final medical decisions remain under the authority of human physicians.

The diagnostic and treatment reasoning of top-tier tertiary hospital experts can be replicated, passed on, and deployed remotely. At a time when the shortage of general practitioners at the primary care level urgently needs to be addressed, the complementarity between AI doctors and real physicians, along with the integration of online and offline services, can meet the multi-level and diverse health needs of the population at the grassroots level.

At the Hudai Town Health Center in Wuxi, Jiangsu Province—a primary healthcare facility serving nearly 100,000 local residents—the Tairex AI Hospital System was introduced for the first time to explore a service model that combines AI doctors with human physicians and integrates online and offline care. In internal medicine and surgical outpatient clinics, the AI-assisted system frees doctors from cumbersome paperwork, allowing for more thorough consultations and physical examinations. Intelligent diagnostic recommendations have also effectively enhanced the standardization and safety of diagnoses. AI has not only become a valuable assistant to physicians but also enabled residents to access better medical services right at their doorstep.

The Tairex AI Hospital System also demonstrates its immense potential in inclusive public health.

In December 2025, the Tiantongyuan North Community Health Service Center in Changping District, Beijing, as one of the first units to undergo public testing of the system, began trial operations serving the 140,000 residents of the Tianbei community. Features such as AI-assisted decision-making and intelligent pre-consultation have helped improve the response efficiency of primary healthcare services and enhance the accessibility of public health services for residents.

“AI is not here to replace us, but to assist us, making diagnosis more standardized and safer.” After practical application, frontline clinicians reported that the system effectively reduced oversights and improved the quality of primary healthcare.

The current market is replete with medical AI products that remain at the stage of standalone tools, where imaging recognition, medical record transcription, and consultation assistants operate as isolated silos, often lacking synergy. At times, these tools face the dilemma of becoming dormant immediately after deployment, resulting in low utilization frequency within clinical diagnosis and treatment workflows.

In particular, many medical AI products may possess medical knowledge but lack an understanding of hospital operations. For instance, they can answer questions and write medical records, yet they do not know how to integrate with hospital information systems (HIS), comply with health insurance regulations, or adapt to the fast-paced workflow of physicians who see dozens of patients per day.

To integrate into the healthcare system, a key requirement is understanding how hospitals operate.

Based on this, the Tairex AI Hospital System focuses on comprehensive systemic enhancement, with AI intervening as early as the pre-consultation stage to help identify and address any omissions. To achieve this, Tairex has deeply integrated into the clinical workflow, covering key stages such as medical orders, examinations, and follow-ups. Meanwhile, AI functionalities are seamlessly incorporated into physicians’ existing operational habits and workflows, requiring no changes to the original system usage logic, thereby minimizing barriers to clinical adoption and implementation resistance.

Compared to traditional diagnostic and treatment workflows, AI-native diagnostic and treatment systems enable closed-loop management across both in-hospital and out-of-hospital settings.

On June 30, 2025, Tairex released the Tairex AI Hospital System. The system features three client terminals designed for patients, physicians, and hospitals, covering the entire care journey including pre-consultation, triage, preliminary inquiry, consultation, and post-consultation phases. Evolvable agent technology has been deeply integrated into every stage, connecting intelligent services both within and outside the hospital.

It is understood that the system achieves low-cost AI empowerment and resource integration for existing physical hospitals by connecting to offline service points distributed across multiple regions.

This can generate substantial value for doctors, patients, and hospitals alike. For physicians, the Tairex AI Hospital System handles labor-intensive tasks such as pre-consultation triage and automated medical record documentation, thereby liberating them from cumbersome paperwork and allowing them to refocus on their core roles: making complex clinical decisions and providing humanistic care.

Especially on the patient side, traditional diagnosis and treatment operate as an “open-loop” system. Patients register, consult, receive prescriptions, and leave… This leads to a heightened risk of breaks in the data chain concerning out-of-hospital recovery, medication adherence, and recurrence risk. The Tairex AI Hospital System provides AI-powered general health consultations, intelligent and precise triage, lifelong exclusive health records, and automated post-consultation follow-ups, delivering proactive, personalized health management services that cover the entire lifecycle.

In terms of data security and privacy protection, Tairex has obtained the Level 3 Certification for Information Security Multi-Level Protection Scheme (MLPS), achieved full ISO privacy and security certifications, and implemented stringent measures—including data encryption, access control, and operational auditing—to ensure security.

To build an AI-driven closed-loop healthcare system, the Tairex AI Hospital System is evolving into a foundational infrastructure that enables continuous tool evolution and seamless data closed-loop circulation.

From the perspectives of regulation and industrial synergy, this path is only just beginning. Nevertheless, Tairex remains confident: “By continuously enhancing our capabilities, we anticipate that industry and policy will converge when market opportunities open up or certain standards are met.”