Scholarlink

<p>The fragmentation of information flows and responsibilities in healthcare generates significant inefficiencies and high coordination efforts for patients. AI-powered agents connected to heterogeneous resource systems via standardized protocols such as the Model Context Protocol (MCP) offer a&#xa0;promising coordination layer. However, it remains unclear which agentic (based on autonomous, goal-oriented AI agents) architecture—centralized single-agent or decentralized multi-agent—is better suited to the complex demands of care coordination. This paper employs a&#xa0;Design Science Research approach: two competing prototypes are developed on a&#xa0;shared MCP infrastructure and evaluated against ten systematically constructed test cases. Results show that single-agent systems outperform in atomic, low-tool-density tasks, while multi-agent systems generally excel in complex, cross-domain, and safety-critical scenarios. A&#xa0;key architecture-independent finding is the persistence problem of safety-critical context: LLM-native knowledge of contraindications and risk factors fades over longer conversations regardless of architectural choice. From the comparative evaluation, four design principles are derived: domain-based specialization, deterministic safety-state, explicit handover schemas, and a&#xa0;complexity-adapted architecture threshold. The contribution provides prescriptive design knowledge for building robust, scalable, and auditable agent architectures for healthcare.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Gestaltung agentischer KI-Systeme für die digitale Versorgungssteuerung: Ein Design-Science-Research-Ansatz

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月6日

查看全文
  • Björn-Lennart Eger,
  • Laura Biller,
  • Daniel Rose,
  • Florian Bontrup,
  • Barbara Dinter,
  • Christoph Kollwitz

摘要

The fragmentation of information flows and responsibilities in healthcare generates significant inefficiencies and high coordination efforts for patients. AI-powered agents connected to heterogeneous resource systems via standardized protocols such as the Model Context Protocol (MCP) offer a promising coordination layer. However, it remains unclear which agentic (based on autonomous, goal-oriented AI agents) architecture—centralized single-agent or decentralized multi-agent—is better suited to the complex demands of care coordination. This paper employs a Design Science Research approach: two competing prototypes are developed on a shared MCP infrastructure and evaluated against ten systematically constructed test cases. Results show that single-agent systems outperform in atomic, low-tool-density tasks, while multi-agent systems generally excel in complex, cross-domain, and safety-critical scenarios. A key architecture-independent finding is the persistence problem of safety-critical context: LLM-native knowledge of contraindications and risk factors fades over longer conversations regardless of architectural choice. From the comparative evaluation, four design principles are derived: domain-based specialization, deterministic safety-state, explicit handover schemas, and a complexity-adapted architecture threshold. The contribution provides prescriptive design knowledge for building robust, scalable, and auditable agent architectures for healthcare.