Skip to main content
QUICK REVIEW

[论文解读] Human Tool: An MCP-Style Framework for Human-Agent Collaboration

Yuanrong Tang, Peng, Huiling|arXiv (Cornell University)|Feb 13, 2026
Human-Automation Interaction and Safety被引用 0
一句话总结

论文提出了 Human Tool,一种 MCP 风格的抽象,将人类视为在 AI 主导的工作流中可调用的工具,从而提升在 AI 优势任务中的性能并降低工作量。控制研究显示相较于纯 AI 基线,结果更好、协作更均衡。

ABSTRACT

Human-AI collaboration faces growing challenges as AI systems increasingly outperform humans on complex tasks, while humans remain responsible for orchestration, validation, and decision oversight. To address this imbalance, we introduce Human Tool, an MCP-style interface abstraction, building on recent Model Context Protocol designs, that exposes humans as callable tools within AI-led, proactive workflows. Here, "tool" denotes a coordination abstraction, not a reduction of human authority or responsibility. Building on LLM-based agent architectures, we operationalize Human Tool by modeling human contributions through structured tool schemas of capabilities, information, and authority. These schemas enable agents to dynamically invoke human input based on relative strengths and reintegrate it through efficient, natural interaction protocols. We validate the framework through controlled studies in both decision-making and creative tasks, demonstrating improved task performance, reduced human workload, and more balanced collaboration dynamics compared to baseline systems. Finally, we discuss implications for human-centered AI design, highlighting how MCP-style human tools enable strong AI leadership while amplifying uniquely human strengths.

研究动机与目标

  • 推动在 AI 优势任务中将编排从人类转向 AI 的必要性,以减少协作瓶颈。
  • 定义一个具备能力、信息与权限的结构化人类工具抽象,供 AI 调用人类输入。
  • 将 Human Tool 实现为带有工具模式和调用协议的 MCP 风格接口。
  • 通过受控实验证明 Human Tool 能提升任务性能并降低感知工作量。

提出的方法

  • 将 Human Tool 定义为一个结构化、可调用的抽象,代表以三维度:能力、信息、权限 的人类贡献者。
  • 通过分层任务分析和三种调用条件来确定何时调用 Human Tool:能力互补、信息交换、权限控制。
  • 建立与人类的沟通模式,使用互动行为和轻量级准则以最小化协作开销。
  • 将框架实现为 MCP 风格接口,后端用 Python(LangGraph)、GPT-4o 推理、MySQL 存储,前端用 React TypeScript 实现结构化编排。
  • 在两个任务域(旅行规划和故事写作)中进行受控实验,将 Human Tool 与 AI Tool 基线进行比较,采用标准化可用性与工作量量表及客观任务结果。
Figure 1 : Contrasting paradigms of human-AI collaboration: Human Tool versus AI Tool. In the Human Tool paradigm, humans are exposed to the agent as MCP-style callable interfaces rather than workflow leaders.
Figure 1 : Contrasting paradigms of human-AI collaboration: Human Tool versus AI Tool. In the Human Tool paradigm, humans are exposed to the agent as MCP-style callable interfaces rather than workflow leaders.

实验结果

研究问题

  • RQ1研究问题1:与 AI 工具基线相比,Human Tool 是否提升任务性能并降低工作量?
  • RQ2研究问题2:该框架是否可在不同任务类型(规划与创意任务)中适应?
  • RQ3研究问题3:Human Tool 是否提升人机协作中的参与度与协作动态?

主要发现

  • Human Tool 在两个任务上均优于 AI Tool 基线,任务准确性更高(旅行规划:86.72% 对 72.66%;故事写作:68.38 对 58.56),胜出率更高(故事写作:0.611 对 0.371)。
  • 参与者在 Human Tool 下报告认知负荷降低、协作满意度提高,认知负荷指标(CSI)均值为 75.48 对 52.83,故事写作中精神努力显著降低。
  • 可用性(SUS)在两个任务中对 Human Tool 更高(旅行规划:70.89 对 58.22;故事写作:79.79 对 60.21)。
  • 调用日志显示在决策关键点需要人类输入,改善了输入的相关性与时机(偏好、责任边界、知识识别)。
  • 定性访谈表明参与者将系统视为伙伴,能够更深入地共同探索、顺畅地共同开发想法。
Figure 2 : An integrated framework for representing Human Tool within AI-managed workflows. It organizes three dimensions: How to define Human Tool, when to call them, and how to communicate effectively.
Figure 2 : An integrated framework for representing Human Tool within AI-managed workflows. It organizes three dimensions: How to define Human Tool, when to call them, and how to communicate effectively.

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。