A framework for longitudinal health AI agents.

To demonstrate the scope and applicability of our framework, we ground it in three representative longitudinal health contexts: (1) chronic symptom management for endometriosis, (2) post-discharge follow-up after heart failure, and (3) ongoing mental health support for anxiety. These cases were selected to reflect distinct yet common patterns in longitudinal health, spanning chronic self-management, transitional care between clinical encounters and sustained personal health support. They represent variation in clinical structure, stability of goals, involvement of care teams and levels of individual responsibility that commonly occur in real-world health trajectories. By examining our framework across these diverse but complementary scenarios, we illustrate how Coherence, Continuity, Adaptation and Agency can support longitudinal alignment under different forms of uncertainty, coordination demands and evolving user needs. Endometriosis is a chronic inflammatory condition characterized by cyclic and non-cyclic pelvic pain, fatigue, gastrointestinal symptoms and fertility concerns that often persist for years before and after diagnosis 47 , 48 . Clinical encounters may be episodic and focused on acute escalation or procedural decisions. As a result, patients frequently engage in extended periods of trial-and-error self-management, attempting to connect fluctuating symptoms with triggers and interventions in the absence of a clear or predictable disease progression 49 . Existing self-management tools, such as symptom-tracking applications, can enable users to log symptoms and relate them with cycle data 50 , 51 . However, hypotheses about triggers, the reasoning behind trying or discontinuing a therapy and the contextual factors shaping a flare require users to continuously reconcile past experiences with current symptoms. Consequently, patients may repeat ineffective strategies or delay necessary adjustments in care 48 , 49 . A longitudinal health agent could address some of these demands by maintaining a coherent, structured understanding of the patient’s condition over time (Coherence) as exemplified by Fig. 2 . Rather than storing only symptom counts or medication lists, the agent preserves prior hypotheses about patterns (‘pain intensifies 2 days before menstruation when sleep is poor’), contextual explanations (‘stressful work weeks amplify fatigue and pelvic discomfort’) and the reasoning behind therapeutic trials (history). When new symptoms arise, the agent situates them within this evolving model, allowing patients to refine their understanding of complex and personally variable patterns without restarting from scratch (persistence). Building on this structured understanding, Continuity ensures that open questions and partial trials are not lost across time. In endometriosis management, interventions often require weeks or months to evaluate, and outcomes may be incremental rather than binary 48 . Therefore, the agent carries forward unresolved threads, such as whether a new hormonal regimen reduced mid-cycle pain, whether dietary changes altered gastrointestinal symptoms, whether worsening fatigue warrants clinical reassessment, and reintroduces them at clinically and personally meaningful moments (follow-up). The agent can then sustain engagement with long-term goals rather than framing progress as discrete task completion (alignment, accountability). In doing so, the agent shares the burden of longitudinal oversight that patients with under-supported chronic conditions often shoulder alone. Because disease expression and life context shift over time, Adaptation is critical. The agent continuously monitors for meaningful changes in symptom intensity, functional impact or life circumstances and adjusts both its recommendations and its underlying assumptions (reflexivity). If a previously effective strategy loses efficacy, the system revises its model rather than attributing divergence to non-adherence. If priorities shift, for example, from pain reduction to fertility planning or vice versa, the agent recalibrates its guidance accordingly (responsiveness). This form of reflexive adaptation acknowledges that both the illness and the patient’s goals evolve, and that responsible support requires periodic re-examination of earlier inferences. Finally, sustained self-management in endometriosis requires careful attention to Agency. Patients often become de facto experts in their own condition, yet may also experience fatigue, dismissal or uncertainty in clinical encounters 52 . A longitudinal agent can modulate its initiative over time, offering more structured guidance during periods of confusion or symptom escalation and shifting towards collaborative reflection as patients gain confidence in recognizing patterns (proactivity, emancipation). By making its reasoning transparent and open to contestation, the agent supports informed self-determination rather than passive compliance (transparency). This calibrated partnership helps preserve both autonomy and safety across the long arc of chronic illness management (negotiation). The transition from hospital to home after heart failure hospitalization is a clinically fragile and operationally complex period 53 , 54 . Unlike conditions with ambiguous trajectories, heart failure management is organized around explicit physiological thresholds and defined recovery milestones, yet successful recovery depends on coordination between patients, clinicians and caregivers 55 . Small lapses in adherence or monitoring can quickly lead to decompensation, making it critical that responsibilities are enacted reliably across multiple actors and settings. Existing post-discharge tools, such as printed instructions and reminder apps, provide guidance but cannot maintain a live understanding of the patient’s recovery trajectory 56 . These tools can signal individual events, like missed medications or rising weight, but they do not interpret patterns relationally or in the context of the original discharge plan 57 – 60 . When care spans hospital, primary care and specialty follow-up, fragmentation arises not from absent guidance, but from the absence of a persistent, integrated view that links emerging data to clinical intent. In this context, Coherence means preserving a live model of the recovery plan and its execution. The agent situates weight changes, medication logs and symptom reports within the context of discharge instructions and clinical thresholds (history). Clinicians can review the same structured timeline, ensuring that the reasoning behind alerts or recommendations is clear (transparency). This shared representation allows the agent to highlight meaningful deviations, such as missed doses combined with rising weight, and helps the patient, agent and clinician see patterns together (relationship). By maintaining explicit links between prescribed strategies and physiological response, the agent reduces the risk that critical signals remain siloed (persistence). Coherence here is less about exploratory sense-making and more about ensuring that the biomedical logic of recovery remains intact as responsibility shifts from hospital to home (organization). Given the time-sensitive nature of post-discharge risk, Continuity can then help track whether medications are taken on schedule, follow-up appointments are completed and symptoms are monitored and re-engage if gaps occur (follow-up) 58 , 59 . Continuity in this setting mediates between clinic and home and preserves feedback loops between the patient’s actions, physiological signals and clinical oversight (alignment). In doing so, the agent ensures that care remains coordinated and coherent across settings and actors rather than fragmented into disconnected reminders (follow-up, accountability). As recovery progresses, Adaptation allows the agent to adjust its support. During the early post-discharge period, the agent may prioritize frequent check-ins and alerts to prevent clinical instability (follow-up). Once the patient stabilizes, the agent can reduce routine check-ins but continue to flag important changes in symptoms or treatment adherence (responsiveness). When clinicians modify treatment or adjust targets, the agent revises its recommendations accordingly, maintaining alignment with evolving clinical guidance (alignment, personalization). This allows the agent to support both patients and care teams as circumstances change (responsiveness). Finally, sustainable management depends on cultivating patient capacity without withdrawing necessary safeguards. Agency is negotiated among the patient, the agent and the clinician (negotiation). In the immediate post-discharge period, the agent may assume a more directive role to reinforce adherence and safeguard against risk (proactivity). Over time, the agent can shift towards collaborative planning by guiding the patient through reviewing symptom and adherence trends, highlighting which behaviours are contributing to recovery, and suggesting when reaching out to a clinician may be warranted based on emerging patterns (proactivity). By making its reasoning transparent and flexible, the agent supports informed decision-making without undermining clinical authority or patient autonomy (transparency). Unlike highly structured medical conditions, mental health trajectories often involve a mix of formal clinical encounters, intermittent therapy, medication adjustments and extended periods of personal health management 61 . Clinical involvement may be sparse, irregular or entirely absent, leaving individuals to interpret subtle changes in sleep, motivation, social engagement or mood on their own 62 . At the same time, when clinicians or therapists are involved, care is structured around specific goals or interventions that must be coordinated with daily coping strategies. Existing digital mental health tools, including both clinically integrated systems and consumer self-help applications, however, are typically designed around isolated interactions. Mood-tracking apps, conversational agents and crisis services capture snapshots of emotional state or intervene in acute moments without visibility into the trajectory that preceded escalation 62 . This intermediate space that can be structured and self-directed creates unique challenges for longitudinal support, where goal stability, individual responsibility and external stakeholder involvement vary over time. Through the Coherence layer, the agent could construct a cumulative and evolving understanding of an individual’s emotional patterns, triggers, coping responses and therapeutic goals. Anxiety and depression are context sensitive and probabilistic with the same behaviour or mood shift signalling different states depending on circumstances 63 . Coherence would allow the agent to differentiate these patterns by connecting current reports to prior episodes, contextual stressors and previously attempted strategies (history). Coherence could also encompass memory of user preferences for engagement, such as how the user responds to prompts, what forms of validation are helpful and when encouragement or gentle guidance is appropriate (relationship). If clinicians are involved, the agent can represent their guidance alongside self-management strategies, ensuring that both patient and care team share a coherent understanding of progress and priorities (organization, persistence). The Continuity layer could then help maintain presence across the periods between therapy or psychiatric appointments (follow-up). The agent can monitor patterns that historically precede relapse, such as changes in communication, disrupted routines or missed self-care behaviours, and intervene proactively (proactivity). Rather than issuing generic prompts, it could situate current difficulties within prior cycles of recovery and relapse, reinforcing that temporary setbacks fit within a broader trajectory (alignment). Continuity in this domain serves both preventive and stabilizing functions by sustaining engagement during stable periods and enabling earlier recognition of destabilization during vulnerable ones (accountability, responsiveness). Users’ emotional states can shift rapidly, and interventions that are helpful in one phase may feel burdensome or misaligned in another. The Adaptation layer could recalibrate tone, structure and intensity of support to respond to these shifts in real time (responsiveness). During a period of stability, the agent might emphasize goal setting, skill development or reflective exercises (personalization). During a period of distress, it could simplify interactions, prioritize grounding techniques, or encourage timely contact with clinicians (responsiveness). Because mental health symptoms themselves can impair motivation and executive function, Adaptation involves dynamically adjusting expectations in real time with symptom severity, cognitive capacity, and contextual stressors (alignment). With potential reductions in decision-making capacity, making some forms of choice or responsibility can be overwhelming. However, complete removal of agency can foster helplessness 64 . The Agency layer in a longitudinal health agent negotiates this balance, dynamically calibrating the locus of control (negotiation). This negotiation can be informed by longitudinal context, allowing the agent to track patterns of decision-making capacity over time, and observe when the individual reliably responds to prompts, initiates coping behaviours independently or benefits from structured scaffolding (reflexivity). The agent could strategically adjust who drives action and provide support that aligns with both immediate capability and long-term self-management goals (proactivity, emancipation).

To operationalize longitudinal support in AI-driven health agents, we synthesized insights from longitudinal clinical practice and personal health informatics, drawing on approximately 40 prior studies spanning human–computer interaction (HCI), personal health informatics and clinical care. This framework captures the functional and relational requirements for maintaining meaningful longitudinal engagement and supporting evolving health goals. Our framework has four interdependent layers: Coherence, Continuity, Adaptation and Agency ( Fig. 1 ). Within each layer, dimensions describe how agents can translate principles into practice. The framework emerged iteratively through extensive discussion and reflection within our interdisciplinary team, including experts in human–computer interaction, AI, health informatics and clinical care. These conversations helped us refine the framework’s scope, make assumptions clear and highlight practical design insights, ensuring it is both well grounded and useful for supporting longitudinal health engagement. Throughout this paper, we use the term ‘agent’ to refer to the system providing longitudinal support, as the appropriate system architecture and its associated trade-offs remain an open design question. In practice, these capabilities may be implemented within a single system or distributed across multiple coordinated agents (for example, a multi-agent system in which each layer is associated with a distinct agent). Classic continuity-of-care frameworks emphasize that care should be experienced as connected and coherent over time. These frameworks highlight the importance of attending to evolving patient priorities, understanding relationships between triggers and symptoms, and assessing the effectiveness of past interventions as part of longitudinal support 22 , 23 . Achieving this requires more than merely storing historical data; it entails reconciling longitudinal personal health management practices and evolving the relationships among patients, clinicians and other health stakeholders, recognizing their respective responsibilities, implementing consistent management strategies, and leveraging insights from prior encounters to guide current decisions 24 , 25 . Current health informatics systems and personalized health agents often do not fully align with these continuity-of-care principles. Many current health tracking systems and LLM-based agents operationalize longitudinal support by storing and reusing accumulated data 11 , 12 . For instance, emerging LLM-based agents with multisession capabilities often summarize or embed prior interactions, then reintroduce them at inference time to improve response relevance 26 – 28 . However, these systems are not designed to explicitly link individual interactions to a user’s broader goals or intentions. Therefore, they cannot currently characterize the evolving relationships among management strategies, symptom patterns, behavioural triggers and subsequent health outcomes. In health contexts, where progress depends on understanding how these factors interact and change over time, this limitation reduces the system’s ability to meaningfully connect a user’s experiences, decisions and outcomes across sessions. When systems fail to explicitly capture user intent and data relationships, critical elements of longitudinal support are lost. Prior assessments may not influence future decisions, evolving intervention–outcome relationships go untracked, and context-dependent priorities are overlooked. As a result, reasoning can break down between interactions, and the system struggles to maintain a coherent and accurate narrative of an individual’s health over time. Our framework addresses these gaps by reconceptualizing Coherence as active, structured sense-making rather than passive memory. Historical data are not treated as a fixed input. Instead, the Coherence layer captures interpretations, hypotheses and reasonings that link past experiences, including both between-visit interactions and in-visit encounters, to ongoing care. We define Coherence along four interrelated dimensions that collectively reinforce stability across interactions as shown in Table 1 . Effective longitudinal care relies on Coherence to maintain a structured understanding of past interactions and relationships. Complementarily, Continuity ensures that this understanding guides ongoing action, goals and follow-up across encounters. Clinical research defines continuity as the sustained and coordinated management of a patient’s condition across encounters 23 , 24 . This body of research emphasizes follow-up, responsiveness to evolving needs, prevention of fragmentation and ‘safety-netting’ when symptoms persist or uncertainty remains 29 – 31 . Chronic care frameworks and proactive-care frameworks underscore systematic reassessment, planned follow-up and structured management of unresolved concerns, particularly for long-term conditions such as diabetes or depression 29 , 31 – 33 . These approaches highlight that meaningful progress depends on maintaining attention to goals that extend beyond a single visit while preserving safeguards against overlooked or worsening conditions. Current digital health systems address part of this need through reminders, task lists and Just-In-Time Adaptive Interventions 34 – 39 . Such systems are effective for prompting discrete behaviours (for example, medication adherence, appointment attendance, daily logging) and improving short-term compliance 36 , 38 . However, most existing systems do not represent partial progress, revisit goals without resetting them or acknowledge forward movement when outcomes are incremental or episodic. Continuity in these systems is often instantiated as timely prompts for predefined tasks, rather than as sustained engagement with evolving and sometimes ambiguous goals. Many health trajectories do not follow a linear path as symptoms may fluctuate, interventions can yield partial or delayed effects and progress may be non-linear. In these cases, continuity is less about completing tasks and more about maintaining orientation towards a long-term aim despite uncertainty while ensuring that unresolved issues remain visible and monitored. Our framework reframes continuity in agents as actively sustaining momentum across interactions. Agents identify unresolved threads, revisit them constructively and maintain alignment between short-term actions and broader health intentions. Instead of treating goals as binary states (completed versus incomplete), the agent tracks their trajectory and how they evolve, stall or transform over time, and supports users in re-engaging without fragmentation or loss of context. We operationalize Continuity through three interrelated dimensions as shown in Table 2 . In addition to ongoing management and oversight of goals (Continuity), longitudinal support requires recalibration of agent assumptions and behaviour as health goals, circumstances and user behaviour transition. Prior work emphasizes that effective health tracking tools must capture and respond to changes in users’ internal and external conditions 29 , 40 , 41 . Proactive and chronic care frameworks highlight that clinicians anticipate changes in patients’ conditions, tailor interventions accordingly and regularly reassess goals to maintain progress and reduce risk 29 , 32 , 33 . Contemporary digital health systems handle certain aspects of this problem. Just-In-Time Adaptive Interventions and reinforcement learning-based health agents, for example, adjust recommendations based on contextual signals, behavioural feedback and short-term engagement metrics to optimize intervention timing or content 15 , 35 – 37 , 42 . However, these approaches often assume stable goals or reward structures. They adapt actions in response to immediate signals but rarely revisit the higher-level assumptions, inferred preferences or causal models that guide those actions. Systems that account for evolving user circumstances typically rely on user-initiated reflection 34 , 40 , 41 and lack continuous internal representation of priorities, risk tolerance and context-dependent assumptions. Our framework conceptualizes adaptation as structured recalibration across time and care contexts, extending beyond reactive adjustment. The Adaptation layer supports not only changes in recommendations but also revision of the interpretive assumptions that guide those recommendations. By pairing personalization with reflexivity, we introduce the notion of second-order adaptation where the agent can revisit earlier inferences about user preferences, goals or risk tolerance and renegotiate them when emerging evidence suggests misalignment. Adaptation should also consider factors beyond the user to account for external changes that can substantially impact users’ health journeys and interactions with the agent, such as evolving clinical guidelines, treatment standards or regulatory updates. We define Adaptation through three interrelated dimensions as shown in Table 3 . Longitudinal care often involves shifting roles, responsibilities and authority between patients, clinicians and supporting systems over time. Clinical frameworks on shared decision-making, patient activation and collaborative care emphasize that effective care depends on explicit negotiation of responsibility 29 , 32 . Clinicians guide and intervene when needed, particularly when safety or clinical risk is present, while patients retain meaningful participation in decisions that affect their lives. Informatics research similarly posits that intelligent systems should not only offer timely support but also respect users’ decision-making capacity and right to self-determination 34 , 40 , 41 . Existing health agents partially engage this tension, but often implicitly. System outputs can be tailored to stated preferences or past behaviour, and users may accept, reject or ignore recommendations to preserve autonomy 43 , 44 . Research on recommender systems and mixed-initiative interaction often frames agency as a moment-level choice between system suggestion and user override 45 , 46 . While valuable, this framing treats agency and autonomy as a transactional property of individual interactions rather than a longitudinal construct shaped cumulatively over time. Persistent deference may overburden users during periods of vulnerability or fatigue, while persistent direction or overly affirming or ‘sycophantic’ AI interaction patterns may gradually erode autonomy and encourage excessive dependence 41 , 45 . Existing systems rarely model or manage these cumulative effects of guidance across sessions. Our framework reframes agency as longitudinal calibration of authority and responsibility rather than one-off consent or momentary interaction preferences. The Agency layer enables the agent to intentionally adjust its degree of initiative, guidance and intervention in response to evolving user capacity, context and goals. In this way, the agent can sometimes take a very active role, such as coordinating with care teams or facilitating communication with external professionals (for example, briefing care teams at the start of visits and integrating visit notes into memory to support Coherence and Continuity). At other times, it may present itself more as a support tool, empowering the user to explore options and make decisions independently. Over time, rather than assuming a fixed balance between system control and user choice, the agent treats agency as something negotiated and revisited across time. We conceptualize Agency across four interrelated dimensions as shown in Table 4 .

LLM-based health agents are increasingly proposed as tools to support complex and evolving care processes. However, most current systems remain episodic, reactive and limited to discrete tasks. In this Perspective, we argue that meaningful health support requires more than accurate responses within isolated encounters. It depends on structured, longitudinal stewardship of goals, interpretations and responsibilities over time. Drawing on insights from continuity of care and personal health informatics, we introduce a four-layer framework that conceptualizes longitudinal support as an active design challenge rather than a by-product of memory or personalization. This framework shifts attention away from static data retention and short-term adherence towards sustained intent, evolving interpretation, negotiated authority and accountable follow-up across interactions. The use cases demonstrate both the potential and the complexity of this approach. Longitudinal agents may enhance coherence, engagement and personalization, but they also raise new concerns related to safety, calibration and the appropriate delegation of responsibility. Addressing these challenges will require new evaluation paradigms, system architectures and interface designs that explicitly support ongoing, multisession engagement, rather than treating continuity as an extension of single encounters.

This section outlines key tensions, challenges and open questions that arise when designing and evaluating longitudinal health agents. Below we highlight core design tensions that emerge when balancing competing longitudinal properties within agent systems. Our framework reveals a tension between information stability and necessary revision through its Coherence layer and Adaptation layers. Overemphasizing coherence risks solidifying tentative assumptions into misleading long-term beliefs, whereas prioritizing adaptation too heavily can fragment the user’s narrative. This tension could be worsened by the risk that initially reasonable outputs may become problematic if treated as persistent authority without reassessment, especially as external knowledge (for example, clinical guidelines and regulation) evolves. Agents must therefore track not just relevant information but the status and basis of interpretations to distinguish durable knowledge from revisable hypotheses and incorporating oversight or verification when needed. These challenges can complicate evaluation. While storage is easy to measure, it is far harder to assess whether systems retain and revise the right information over time, particularly as failures may stem from outdated knowledge, flawed reasoning, or appropriate uncertainty. Ultimately, ensuring alignment requires longitudinal evaluation of coherence drift and user perceptions of whether revisions are justified, transparent and responsive to changing evidence. A parallel tension exists between short-term task execution and long-term trajectory. Discrete tasks, such as trying a new intervention, monitoring symptoms or scheduling follow-up care, remain essential, yet their impact is often realized only through accumulation, comparison and reflection across extended periods. Current evaluations of health agents tend to prioritize diagnostic accuracy or task completion within single interactions 9 , 65 , but this narrow focus can misclassify pauses or regressions as failure, despite non-linear progress being typical in health contexts 10 . Supporting Continuity therefore requires representing goals as evolving pathways rather than binary endpoints. Agents must help users see how incremental effort, reflection and recalibration contribute to a broader direction, even when outcomes remain uncertain. Evaluating this capability cannot rely solely on short-term adherence metrics alone, as systems might optimize isolated task completion without sustaining engagement with long-term goals. Demonstrating true continuity may require following users across extended periods to examine whether unresolved concerns are revisited, whether partial gains are acknowledged and built upon, and whether users maintain orientation towards meaningful objectives despite interruptions. Treating Agency as a longitudinal variable introduces a tension between proactive support and respect for autonomy, as independence is not fixed but shifts with users’ conditions, circumstances and goals. Rather than a linear progression towards full independence, agency becomes an ongoing negotiation of roles and authority, requiring systems to make their level of initiative both legible and adjustable. Agents must balance when to intervene and when to step back, as persistent over-direction can undermine confidence while excessive deference may leave risks unaddressed. Of note, autonomy may also include the user’s ability to retract or delete previously shared data, which can reshape the agent’s contextual understanding and limit its capacity to provide longitudinally informed support. This underscores the need for longitudinal systems to handle such missing data transparently, recalibrate appropriately and communicate any resulting uncertainty or limitations in guidance. These cumulative effects are difficult to capture in traditional evaluation paradigms. Although current personalization methods often rely on moment-level behavioural data like adherence or engagement 42 , such metrics do not reveal whether users become more capable, dependent or disengaged over time. Assessing longitudinal agency instead requires combining behavioural data with repeated measures of perceived competence, trust and autonomy, as well as examining whether proactivity improves safety without diminishing user control, including how other stakeholders perceive the agent’s involvement in care. Emerging agent frameworks already include useful building blocks for longitudinal health agents such as persistent memory, evolving system state and proactive behaviours like periodic triggers and background processes 27 , 28 , 66 . However, these capabilities alone are not sufficient to develop a longitudinal health agent as they do not track changing health goals, separate tentative ideas from stable knowledge and calibrate their control around user circumstances over time. Addressing these will likely require more structured representations of longitudinal state, such as maintaining time-indexed goal hierarchies, explicitly labelled hypotheses (for example, tentative versus confirmed) and relational links between symptoms, interventions and outcomes. In parallel, systems will need mechanisms to ground outputs in verified medical knowledge and to maintain calibrated estimates of uncertainty and clear provenance trails as these internal representations evolve over time. In practice, these systems will also likely not operate in isolation. Longitudinal agents must coordinate across components and stakeholders, whether implemented as multi-agent systems or integrated into existing healthcare infrastructure. This includes interfacing with electronic health records to incorporate validated clinical data and communicate longitudinal insights back to care teams. Such integration raises challenges in maintaining shared representations, avoiding fragmentation and ensuring that reasoning remains coherent across systems, encounters and stakeholders. Differences between patient-reported data and clinical records, delays in updates and mismatched data formats can all introduce inconsistencies that grow over time if they are not actively managed. On the other hand, the system might not always be able to rely on clinician-mediated validation or communication as users might engage in independent self-management practices irrespective of clinical access. In these settings, longitudinal health agents might be more focused on supporting individuals in forming, revising and maintaining their own understanding of their health trajectories over time. This shifts key design requirements beyond coordination with external stakeholders towards a more demanding need for internal mechanisms of self-calibration, including robust communication of uncertainty, safeguards against over-reliance on system outputs and support for situating short-term experiences within longer-term patterns. The accumulation of sensitive health data over time raises important privacy and governance concerns. Longitudinal agents must support ongoing consent, clear data ownership and alignment with regulatory frameworks (for example, HIPAA, GDPR). This may necessitate features for selective memory revision or deletion. As agents adopt more proactive roles, however, their actions may end up depending on incomplete or outdated data, introducing substantial safety, ethical and clinical risks. Thus, ensuring safe deployment will likely require oversight, auditing and governance mechanisms such as human-in-the-loop oversight, auditable reasoning and clear boundaries between assistive support and clinical decision-making. While the allocation of responsibility and liability will depend on evolving regulatory and legal frameworks across jurisdictions, a near-term design priority is for longitudinal agents acting on behalf of users to be explicit about when, why and under whose authority they act, and to retain sufficient context to enable subsequent review and audit. We see this as a shared challenge for designers, developers, clinicians and policymakers to navigate collectively as these systems mature. In Table 5 below, we outline open questions for future research.