Volume 3, Issue 3

Medical artificial intelligence is moving beyond prediction and content generation toward goal-directed reasoning, tool use, and controlled execution, giving rise to new scientific questions that transcend model performance evaluation. We propose agentic medicine science, a discipline that investigates how medical agents participate in clinical action within defined objectives, human oversight, and institutional boundaries, and how such participation affects patient outcomes, care processes, and health system performance. Its fundamental unit of analysis is the patient – professional – agent – health system (PPAHS) composite system. Its  knowledge architecture comprises three pillars: clinical action, clinical evidence, and accountability and governance; global translation is guided by a 4S agenda of standards, science, safety, and solidarity. We distinguish medical GPTs, automated workflows, medical agents, multi-agent systems, and embodied agents, and propose a tiered evidence pathway extending from technical validation and clinical validation to clinical outcomes research and continuous post-deployment monitoring. Chinese practices, including closed-loop pulmonary nodule management and hospital-at-home care coordinated with cloud-based outpatient services, suggest that the value of medical agents lies not in replicating physicians, but in reorganising fragmented knowledge, tasks, and responsibilities into a  continuous care system that is verifiable, controllable, and accountable.

Key Words: medical agents; agentic medicine; large language models; clinical evaluation; artificial intelligence governance; human–AI collaboration; health-system reform

Acute respiratory distress syndrome (ARDS) is a common clinical syndrome associated with high mortality in critical care medicine, characterized by marked intrapulmonary spatial heterogeneity; static indices derived from laboratory tests, respiratory mechanics, and conventional imaging cannot continuously capture the dynamic physiological process of regional ventilation-perfusion mismatch. Although electrical impedance tomography (EIT) offers noninvasive, radiation-free, continuous bedside visualization of regional ventilation distribution, a single modality alone cannot fully reflect the dynamic evolution of disease treatment and prognosis. In recent years, medical artificial intelligence has been evolving from medical GPT (generative pre-trained transformer), which is centered on language generation, toward medical agents that integrate multisource perception, rule-based governance, interpretable reasoning, and human-in-the-loop mechanisms. The Argus agent, jointly developed by Zhongshan Hospital, Fudan University and Suzhou Kaonter Medical Technology Co., Ltd., has implemented multisource device connectivity, an ARDS overview dashboard, continuous EIT video and derived-index display, white-box alerts, and a cloud/local report pipeline, providing a medical-agent paradigm for the clinical implementation of spatiotemporal multimodal monitoring. Our analysis suggests that the value of the Argus agent lies not in 'giving automatic answers' but in revealing regional differences, organizing multisource evidence, and supporting explainable decisions. In the future, with the advancement of standardized acquisition, multicenter studies, and lifecycle governance, the Argus agent is expected to evolve from a 'foundational framework for a monitoring-oriented digital twin' into a trustworthy monitoring-oriented digital twin for ARDS.

Key words: medical GPT; medical agent; electrical impedance tomography; acute respiratory distress syndrome; multimodal monitoring; digital twin

Clinical care requires iterative evidence acquisition, revision of working hypotheses, and action under incomplete information, changing patient states, and finite resources. Yet medical artificial intelligence (AI) is still evaluated largely through recognition, prediction, content generation, or static question answering. The central scientific question is whether model capability can be translated into safe, effective, and accountable longitudinal clinical action. A medical agent is not merely a large language model (LLM), but a sociotechnical system integrating models, knowledge, data, memory, tools, workflows, human oversight, and governance. Drawing on representative studies, authoritative reviews, reporting guidance, and regulatory documents available through July 2026, this Review proposes a six-module closed-loop architecture—perception, knowledge, reasoning, planning, action, and feedback—clarifies boundaries with related AI systems, and links clinical use cases to evidence and governance requirements. Retrieval-augmented generation, tool use, reflection, and multi-agent collaboration have improved performance in selected controlled tasks. However, static medical question-answering scores are weak proxies for interactive clinical performance, and recent benchmarks show only modest accuracy gains despite substantial increases in token use and latency. A small number of randomized trials and real-world workflow studies suggest potential benefits for documentation, trial prescreening, and bounded decision processes, but do not support unsupervised autonomous diagnosis or treatment. We therefore propose four risk-calibrated autonomy levels, a six-stage evidence pathway, and lifecycle governance. Clinical value should be judged by the net benefit of the clinician–agent–workflow system. The near-term goal is not clinician-independent autonomy, but clinician-led, patient-engaged, evidence-constrained, and accountable human–agent care.

Key Words: medical AI agent; large language model; multi-agent system; clinical translation; human–AI collaboration; trustworthy artificial intelligence; lifecycle governance

Medical artificial intelligence is evolving from single-output models into agents capable of maintaining contextual coherence, performing multi-step planning, invoking tools, and initiating or drafting clinical actions. Compared with traditional models, the risk of agents is no longer measured solely by whether a single output is correct; it depends more on how they perform task planning, invoke permissions, write to memory, select tools, and coordinate with the clinical team. Once an error propagates through any of these links, it may translate from a deviation at the level of recommendation into consequences at the level of action. Therefore, evaluation before medical agents enter clinical settings cannot rely only on the model’s own performance metrics; it must cover the complete system formed by the human–agent–workflow. The core question of this paper is: according to the severity of the consequences of medical agent actions, what level of system evidence is required to determine whether an agent can safely enter and continuously operate in clinical settings. Through a narrative review, this study integrates interdisciplinary literature from medical artificial intelligence, human factors engineering, safety engineering, cybersecurity, implementation science, health economics, reporting standards, and regulatory science, and constructs MAAM–MAAE, a risk-adaptive conceptual methodology. We map existing standards to agent-specific attributes and risks and organize unresolved requirements into interconnected R&D and evaluation cycles. It should be noted that this study uses purposive rather than systematic evidence synthesis and has not conducted a formal expert consensus process or prospective validation. Medical Agent Architecture and Maturation (MAAM) divides R&D activities into four domains: clinical mission, system architecture, human–agent collaboration (including action authority), and safety and maturity assurance. At the assessment level, MAAM introduces Medical Agent Assessment and Evaluation (MAAE) to organize evidence across four hierarchical dimensions: capability, safety, human–agent and clinical performance, and health system and full-lifecycle value. The unit of evaluation is the complete human–agent–workflow system. Four risk tiers and five stage gates—the gates are set according to potential harm, the physical proximity of actions to patients, the degree of autonomy, error detectability, and reversibility—so that supervision intensity matches the required level of evidence. In addition, this paper proposes clinical task success (CTS) and clinically consequential unsafe action rate (CCUAR) as dual anchoring metrics; however, their measurement properties and decision thresholds still require empirical testing. MAAM–MAAE transforms clinical authorization from a one-time model validation decision into a version-specific full-lifecycle contract. By incorporating actions, longitudinal operation, change control, and exit mechanisms, this framework complements existing reporting and regulatory frameworks. MAAM–MAAE is a testable proposal, not a validated checklist, reporting guideline, or claim of clinical effectiveness.

Key Words: medical agents; agentic artificial intelligence; clinical evaluation; human–medical agent collaboration; patient safety; full-lifecycle governance; methodology

Key Words: medical agents; trustworthy artificial intelligence; large language models; clinical evaluation; human–machine collaboration; risk stratification; lifecycle governance; patient safety

Screening with low-dose helical computed tomography (LDCT) has substantially reduced lung cancer mortality in high-risk populations, but it has also led to the detection of large numbers of pulmonary nodules that require further risk stratification and long-term follow-up. Existing imaging artificial intelligence (AI) tools have focused primarily on nodule detection or prediction of malignancy risk, and smart reminder tools can reduce missed follow-up in some patients. However, evidence from single AI models has not yet answered whether multi-module medical agents can improve patient-level follow-up outcomes. Therefore, this study aims to determine whether a medical agent equipped with six capabilities—perception, comprehension, prediction, decision-making, execution, and learning—can improve whole-process follow-up management of pulmonary nodules and patient clinical outcomes. BAIMGPT-PN-RWE is a multicenter, prospective, active-controlled real-world study enrolling patients with newly detected nodules. We plan to enroll 6,000 adults with newly detected pulmonary nodules measuring 4– 30 mm from healthcare institutions at different tiers. The primary comparison is between patients who initiate the BAIMGPT six-element closed-loop follow-up management system within 7 days after time zero and those who initiate routine manual follow-up or an existing Internet of Things (IoT)–assisted management system. The follow-up period is 12 months (extended to 24 months in a subset of the cohort). The primary analysis emulates a target trial: propensity score overlap weighting will be used to estimate the average treatment effect in the overlap population; a difference-in-differences approach will be used to assess the effect of the staggered rollout across centers; and marginal structural models will be constructed to control for time-dependent confounding arising from the actual intensity of system use. The primary outcome is the completion rate of guideline-concordant follow-up management during the 12-month follow-up period, i.e., the proportion of patients in whom risk stratification, follow-up planning, follow-up execution, reporting of follow-up results, and next-step decision-making are all completed within prespecified time windows. Key safety endpoints include missed diagnosis of malignant nodules, excessive invasive examinations of benign nodules, failure to escalate follow-up for high-risk nodules, and follow-up delays attributable to the system itself. This study evaluates, in the real world, the value of a six-capability medical agent for follow-up management of patients with pulmonary nodules. Based on the study data, we will further delineate the conditions under which six-element closed-loop medical agents can be generalized, should be restricted, or should be discontinued.

Key Words: pulmonary nodules; medical agents; closed-loop management; real-world evidence; target trial emulation; clinical artificial intelligence

The challenge facing medical research is not only information overload but also the disconnects among evidence, questions, research, implementation, and replication. Research agents can call external tools, integrate multi-source information, and perform multistep tasks, but they may likewise amplify fabricated citations, confirmation bias, version drift, and unauthorized actions into systemic research errors. Existing evaluations of medical artificial intelligence have mostly focused on simulated tasks, and rigorously defined agents remain understudied; there is also a lack of unified definitions, comparable endpoints, and prospective validation pathways for treating research agents as complex research interventions. Accordingly, this article proposes a falsifiable proposition: research agents should be evaluated under prespecified research quality standards by whether they increase reproducible net research benefit without increasing the risk of serious research errors. TRACE-Med organizes the workflow into five stages—Track (evidence tracing), Refine (question refinement), Architect (research architecting), Coordinate (coordinated implementation), and Evaluate (validation and replication)—and uses evidence provenance, prespecification, separation of privileges, and independent reruns as putative mechanisms of action. Clinical implications: This article positions the minimum implementable research unit, Valid and Reproducible Research Gain (VRRG), and Agent-Induced Research Error Rate (AIRER) as candidate components for future evaluation of the implementation effects of research agents, rather than as validated scales or efficacy endpoints. Current evidence supports controlled augmentation, not unsupervised replacement; next steps should employ blinded outcome adjudication, randomized or crossover task designs, multicenter implementation studies, and independent replication to test whether the framework produces net benefit.

Key Words: research agents; clinical research; reproducibility; artificial intelligence governance; open science; TRACE-Med

The key issue in using generative artificial intelligence in medical education (AI-on) is not whether models can complete tasks, but whether learners can retain and transfer clinical competence after AI is withdrawn (AI-off), while recognizing and controlling AI-induced unsafe reasoning. However, existing studies have focused mainly on AI-enabled immediate knowledge, skills, and satisfaction, whereas reports on AI-off transfer ability and changes in patient-related outcomes remain scarce. Therefore, based on narrative evidence syntheses published before July 2026, this article proposes the dual-competency agentic medical education (DC-AME) framework. The framework consists of two complementary competencies: AI-off independent clinical competence and AI-on safe human–AI collaboration competence, organized into a closed loop comprising learner diagnosis, contextual simulation, scaffolding fading, and transfer validation. Built on shared cases, competency objectives, and an evidence base, intelligent tutors, virtual patients, and clinical reasoning agents jointly support a minimal training unit of “AI-off independent answering—AI-on assisted reflection—AI-off transfer to unseen cases.” Nevertheless, DC-AME remains a design proposition awaiting validation. Future studies may designate delayed transfer after AI-off and the recognition and control of AI-induced unsafe reasoning as co-primary outcomes, and list AI reliance, transfer gap, and performance after scaffold withdrawal as exploratory indicators; ultimately, the value of DC-AME should be evaluated in practice.

Medical artificial intelligence is evolving from single-task prediction models and generative assistants into agents capable of perceiving the environment, decomposing goals, invoking tools, coordinating resources, and tracking outcomes. However, a leap in capability does not equate to a leap in clinical authorization. Existing evidence comes mainly from retrospective data and simulated cases, while real-world, multicenter studies with patient outcomes as endpoints remain scarce. If autonomy is pursued without boundaries, accountability, and operational monitoring, agents may amplify a single model error into continuous harm across processes and institutions. This paper proposes the agentic health system (AHS) framework: under the premises of a clearly defined intended use, action authority, accountable entities, and failure fallback, agents execute a “perceive–prioritize–coordinate–act–learn” loop around patient goals, connecting the three domains of hospital, home, and public health, and shifting healthcare from fragmented services to continuous coordination. This paper further proposes four levels of action authority, a minimum implementation unit, a four-stage translation pathway, two anchor endpoints—system safety and system benefit—and the SCORE evaluation framework. The maturity criterion for AHS is not the number of automated tasks, but whether appropriate care can be delivered at the appropriate time, in the appropriate place, and under appropriate supervision, while improving safety, continuity, operational value, resilience, and equity. The endpoint of hospital agents should not be the “unmanned hospital,” but rather a supervised autonomous healthcare system that is professional-led, patient-engaged, machine-auditable, and accountability-traceable.

Medical artificial intelligence (AI) is evolving from algorithms that perform isolated predictive tasks into medical agents capable of continuously perceiving clinical context, decomposing goals, invoking tools, coordinating communication, and tracking outcomes. This transition shifts the central clinical question from whether AI can replace physicians to how patients, physicians, agents, and healthcare institutions can jointly form judgments, authorise actions, and assume responsibility for their consequences. Existing evidence indicates that large language models (LLMs) can perform strongly in medical question answering, differential diagnosis, and management reasoning. Yet “a model outperforms a human” does not mean that “a human–AI team outperforms a human alone”: randomised trials and imaging studies have exposed systemic risks arising from interaction design, automation bias, error amplification, diffusion of responsibility, and erosion of professional competence. We propose the Symbiotic Clinical Compact (SCC), comprising four interrelated commitments—shared intelligence, bounded action, patient co-governance, and traceable accountability—to define the rights, authority, and responsibilities of patients, physicians, agents, and healthcare systems. We further introduce the FADA clinical loop—Frame, Analyse, Deliberate, and Authorise—together with a four-stage translational pathway extending from human review to human–AI symbiotic medicine. Human–AI symbiotic gain (HASG) and symbiosis-related harm (SRH) are proposed as dual-anchor evaluation endpoints. The SCC does not grant agents unrestricted authority to practise medicine. Instead, it embeds task-level autonomy within revocable authorisation, risk stratification, continuous monitoring, and institutional accountability. Future research should move beyond static model validation towards real-world evaluation of integrated human–AI systems, and beyond average accuracy towards patient outcomes, subgroup fairness, relational quality, professional competence, and lifecycle safety. The goal of human–AI symbiotic medicine is to enable agents to manage informational and procedural complexity, physicians to focus on judgment, responsibility, and compassionate care, patients to become genuine partners in medical decision making, and healthcare systems to evolve from episodic service providers into continuously learning, traceable health partners.

Key Words: medical agents; large language models; human–AI collaboration; shared decision making; patient autonomy; clinical governance; traceable accountability; artificial intelligence ethics