Skip to main content
[Preprint]. 2025 Aug 26:2025.08.22.25334232. [Version 1] doi: 10.1101/2025.08.22.25334232

Table 3.

AI Agent Architecture Selection Framework for Clinical Medicine

Architecture Type When to Use (Decision Criteria) Clinical Applications Implementation Requirements Key Considerations & Limitations
Simple Tool Augmentation Direct tool invocation without iteration
  • Single, deterministic operation

  • Well-defined input/output

  • No reasoning required

  • Immediate response needed

  • Clear success metrics

Medication Dosing Lab Value Conversion DICOM HU Extraction ICD-10 Lookup Drug Interaction Check Risk Score Calculation Technical:
  • API integration

  • Minimal compute

  • No model fine-tuning

Clinical:
  • EHR integration

  • Validated algorithms

  • Regulatory clearance for calculations

Limited to discrete tasks No complex reasoning Cannot handle ambiguity
Best for: High-volume, repetitive tasks requiring consistent accuracy
Avoid for: Tasks requiring clinical judgment or context
Single-Agent Systems Autonomous reasoning with tool orchestration
  • Multi-step workflows

  • Tool selection decisions

  • Iterative refinement needed

  • End-to-end task completion

  • Context preservation required

Radiology Reports EMG Interpretation Literature Triage Clinical Summaries Treatment Planning Discharge Notes Technical:
  • ReAct/CoT prompting

  • Memory management

  • Tool API suite

  • GPT-4+ class models

Clinical:
  • Workflow integration

  • Audit trails

  • Human oversight protocols

Context window limits Single perspective Tool timing critical
Best for: Structured workflows with clear objectives
Optimal tools: 3–5 specialized tools
Key success factor: Well-defined task boundaries
Multi-Agent Ecosystems Collaborative specialized agents
  • Cross-specialty

  • expertise

  • Conflicting evidence synthesis

  • Bias mitigation

  • critical

  • Complex consensus needed

  • Interdisciplinary collaboration

Rare Disease Dx Tumor Board Gene Editing Virtual Lab Multi-Omics ICU Management Technical:
  • Agent orchestration

  • Consensus mechanisms

  • Distributed compute

  • Optimal: 4–5 agents

Clinical:
  • Multi-disciplinary protocols

  • Role definitions

  • Conflict resolution

  • Clinical BIases

Coordination overhead Conflicting outputs Error propagation
Best for: Genuinely interdisciplinary challenges
Caution: Diminishing returns >5 agents