关键词: 生成式人工智能, 图书馆系统, 智能体, MCP, A2A

Abstract:

[Purpose/Significance] This paper examines the ongoing transformation of library information systems, shifting from platform-oriented architectures to agent-based ones, in the context of generative artificial intelligence. It argues that, although Integrated Library Systems (ILS) and Library Services Platforms (LSP) have improved workflow automation and resource management, they remain constrained by poor semantic understanding, restricted cross-system orchestration, and insufficient support for proactive, personalized services. Building on these observations, the paper proposes a transformation path in which existing ILS/LSP infrastructures are not discarded, but rather re-positioned as providers of capabilities within a broader ecosystem of generative intelligent agents. This provides libraries facing both legacy constraints and pressures for service innovation with a feasible evolution strategy. [Method/Process] The study first reviews service-level limitations of ILS and LSP through the lenses of interaction patterns, data openness, and intelligent service support, and distills typical pain points encountered in cataloging, circulation, reference services, and subject liaison work. On this basis, it constructs a graded capability model for generative intelligent agents that encompasses semantic perception, context modeling, goal-driven behavior, preference adaptation, and reflective evolution. It also discusses how different types of agents can be aligned with specific library roles and task granularities. The study then proposes a three-layer architecture consisting of a basic service layer, an agent coordination layer, and a semantic interaction layer. The bottom layer exposes atomic capabilities such as search, metadata editing, authentication, and logging; the middle layer orchestrates multiple agents via lightweight protocols and shared task states; and the top layer supports natural-language-driven interaction while maintaining semantic consistency and traceable reasoning paths. Finally, leveraging a "Library Assistant" prototype that integrates these components, the study designs and conducts experimental evaluations in bibliographic follow-up and recommendation scenarios, combining task-based user tests with qualitative feedback from librarians and domain experts. [Results/Conclusions] Experimental results indicate that the proposed architecture outperforms traditional models in terms of answer relevance, interaction fluency, and perceived service intelligence, particularly in multi-step information-seeking and follow-up recommendation tasks. At the same time, the study found that the mechanisms for long-term memory, cross-session user modeling, and explicit feedback loops were underdeveloped. This can lead to inconsistencies in sustained interactions and complex task chains. The paper concludes with a discussion of the design implications for the evolution of library systems, suggesting that future work should focus on trustworthy memory management, transparent agent coordination, and robust evaluation metrics. It also recommends the development of governance frameworks that jointly consider system performance, user experience, professional ethics, and institutional policy requirements together. In this way, the study provides both a conceptual blueprint and empirical evidence to guide the transition from platform-oriented systems to agent-based, generative AI-enabled library architectures.

Key words: generative AI, library information systems, AI agents, MCP, A2A