Clinical Reality vs. Computational Promise: Scoping Review of Agentic AI Systems in Healthcare

Background: Artificial intelligence (AI) systems are transforming healthcare, with single-agent systems offering efficiency for routine tasks and multi-agent systems providing robustness for complex scenarios. However, clinical implementation faces significant barriers related to safety, accountability, equitable access, and significant gaps in clinical validation-particularly for multi-agent paradigms. Methods: We conducted a scoping review following PRISMA-ScR guidelines, synthesizing evidence from 150+ peer-reviewed studies (2018-2024, with select 2025-2026 sources) retrieved from PubMed, IEEE Xplore, and Scopus, supplemented by relevant arXiv preprints. Given substantial heterogeneity in model architectures, clinical tasks, outcome measures, and validation settings across studies, a quantitative meta-analysis was not appropriate. Instead, we employed structured evidence synthesis with GRADE-based certainty assessment and evidence mapping. Our analysis utilizes a pragmatic, deployment-focused definition of "agents" to bridge the gap between computational theory and clinical reality. Results: High-certainty evidence (supported by high-quality evidence, GRADE: High Certainty) shows single-agent systems demonstrate superior speed (seconds to minutes) and efficiency for specialized diagnostics, achieving 94.5% accuracy in retinal screening and radiologist-level pneumonia detection. Very low-certainty simulation evidence (GRADE: Very Low) reports potential for 15% accuracy improvements in multi-agent systems through coordination in computational models, but these gains remain unvalidated in clinical practice. Multi-agent systems require 5-10x higher computational costs and introduce 200-500ms coordination latency in simulation environments. Both paradigms face critical implementation barriers: algorithmic bias affecting minority patients 1.75x more frequently (Moderate-certainty evidence), unclear liability frameworks, and poor integration with clinical workflows (Low-certainty evidence). Current evidence is predominantly from high-income countries (High-certainty evidence), with limited longitudinal outcome data (Very low-certainty evidence). Conclusions: Single-agent systems have achieved clinical validation and production deployment for routine diagnostic tasks. Multi-agent systems remain predominantly experimental with limited clinical evidence; theoretical advantages from simulation studies require rigorous real-world validation before clinical adoption. We emphasize that these findings are based on a pragmatic lens focused on deployment architectures. Urgent research priorities include large-scale clinical trials for multi-agent systems, development of safety frameworks for autonomous medical systems, risk-based regulatory pathways, and mechanisms to ensure global equity. Healthcare systems must establish ethical oversight structures and validate human-AI collaboration models before scaling autonomous systems.