Answers are grounded in this paper and cite their sources.
Introduction. Demand for medical care continues to outpace the supply of clinicians, and AI agents are already conducting patient encounters at scale. The highest-stakes decision an agent makes is whether to route a patient out of self-care and into acute care. We evaluated a dual-agent escalation architecture, deployed at an AI-native physician in the loop virtual care company, in which a dedicated safety agent runs in parallel with a messaging agent, and we assessed how the home-to-acute-care decision should be designed and governed. Methods. The safety agent was evaluated in two settings. In silico, it was benchmarked against four leading foundation models on the consensus subset of a public health benchmark, restricted to single-patient, English-language, escalation-relevant scenarios. In practice, the agentic workflow was deployed to 162 patients recruited by online advertisement between July and August 2025, covering 477 clinical encounters. Two board-certified emergency medicine physicians independently reviewed every conversation and labeled whether they would advise emergency evaluation, using a conservative criterion. The same de-identified encounters were scored by a general-purpose foundation model for comparison. Results. The safety agent flagged 41 of 477 encounters (8.6%) as potential emergencies and achieved 88.0% sensitivity (95% CI 70.0-95.8%) and 95.8% specificity (95% CI 93.5-97.3%) against the physician gold standard, with a positive predictive value of 53.7% and a negative predictive value of 99.3%. The general-purpose model reached 80.0% sensitivity and 44.4% positive predictive value on the same encounters. On the benchmark subset, all models achieved perfect recall, but precision was 0.84 for the safety agent compared with 0.27 to 0.29 for the foundation models. Published results for nurse-led telephone triage and a widely used symptom checker are reported for context only, because those studies used different populations and reference standards. Conclusion. Escalation is the safety-critical center of agentic patient care, and a dedicated safety agent outperformed both a general-purpose model and published comparators on emergency detection. Digital biomarkers, which already inform decisions in heart failure, diabetes, and bipolar disorder, are the next major class of escalation inputs. Phased deployment tied to concordance and safety benchmarks, with explicit rollback criteria, offers a workable governance model for autonomous clinical AI.
No takes yet. Share an insight, caveat, or question.
Whitehead et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: