Transactions on Machine Intelligence 2821-1693 9 2 2026 06 01 Evidence-Gated Autonomy in Healthcare AI Agents: A Systematic Review of Action Scope, Human-Oversight Effectiveness, and Multi-Step Safety Failures 84 111 248202 10.22034/tmi.2026.248202 EN H. Ghayoumizadeh Associate Professor of Biomedical Engineering, Vali-e-Asr University of Rafsanjan 0000-0002-5390-3938 Kh. Rezaee Department of Biomedical Engineering, Meybod University, Meybod, Iran A. Fayazi Department of Engineering.Vali-e-Asr University of Rafsanjan,Iran O. Rahmani Seryasat Assistant Professor,Faculty of Electrical and Computer Engineering, Shams Gonbad Higher Education Institute, Gorgan, Iran 0000-0002-9289-6128 Journal Article 2025 11 23 Background: Healthcare AI agents increasingly perform multi-step workflows and tool use, but evidence on their operational autonomy, safety, and human oversight remains limited. This review examines whether greater autonomy is supported by empirical safety and oversight evidence. Methods: PubMed/MEDLINE, Scopus, and Web of Science were searched from January 2022 to July 2026. Of 4,096 records, 2,557 remained after deduplication. Fifty priority reports were selected for full-text review; 34 were assessed and 28 met the inclusion criteria. Data on workflow, architecture, autonomy, safety, oversight, and evaluation were narratively synthesized. Results: Among 28 studies, 18 were published in 2026. Thirteen addressed diagnostic or clinical decision support, nine EHR or administrative workflows, four treatment planning or prescribing, and two patient-facing or cross-domain applications. Autonomy levels were L0 in five studies, L1 in 13, L2 in three, L3 in one, and L4 in six; none reached L5. Most L4 systems were evaluated only in sandbox or retrospective settings. Only three studies empirically assessed oversight effectiveness. Conclusions: Current evidence supports tool-using and workflow-capable healthcare agents, but not unrestricted clinical autonomy. Greater autonomy should require stronger evidence of safety, failure containment, and effective human oversight.

https://www.tmachineintelligence.ir/article_248202_3b490b92ec85b3e3cb8258a886ca61e9.pdf