Accepted for/Published in: Journal of Medical Internet Research
Date Submitted: Aug 27, 2026
Date Accepted: Oct 13, 2026
Recoverability-Centered Trajectory Governance for Long-Horizon Clinical AI Agents
ABSTRACT
Clinical artificial intelligence (AI) is moving from systems that generate discrete predictions and recommendations toward agents capable of pursuing clinical objectives through repeated tool use, communication, planning, and state-changing actions. Recent systems can navigate electronic health record environments, request investigations, formulate treatments, and initiate permitted clinical operations, while emerging longitudinal-agent frameworks envision sustained support across repeated encounters. This transition changes the fundamental unit of safety. An agent does not merely observe an evolving clinical state; through its actions, it can help create the state on which its subsequent decisions depend. Consequently, individually defensible actions may compose into an unsafe trajectory even when no single action is clearly erroneous. We argue that long-horizon clinical AI requires recoverability-centered trajectory governance. Its central objective is clinical recoverability: preserving enough opportunity for clinicians and patients to redirect an agent-mediated care trajectory before clinically acceptable corrective pathways are exhausted. We introduce the recovery margin as the remaining opportunity for timely and proportionate correction, and reauthorization boundaries as trajectory transitions at which delegated authority should end or narrow pending renewed human judgment. This framing distinguishes reversibility, which is primarily a property of an action, from recoverability, which is a property of an evolving human-AI clinical trajectory. Operationally, trajectory governance requires attention to cumulative action, outstanding obligations, option foreclosure, and declining capacity for rescue. Human oversight should therefore be concentrated at consequential trajectory junctions rather than uniformly distributed across every agent action. Evaluation should move beyond endpoint success toward trajectory stress testing, recovery-margin indicators at escalation, interruption-and-takeover performance, cumulative unreviewed action exposure, and population-level trajectory disparities. This perspective also has regulatory and societal implications. Product lifecycle governance asks whether AI remains safe as technology changes; trajectory governance asks whether continued AI action remains safe as the patient changes. Moreover, the safety of autonomy depends partly on access to timely human rescue, making recoverability an equity issue as well as a technical one. For long-horizon clinical AI, the defining safety question is no longer only whether the agent’s latest action is acceptable, but whether it remains safe to let the agent act again.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.