@article{ZHANG2026, 
author = {Xinxin ZHANG and Huanhuan DU and Yuanyuan XI},
title = {Publishing Agents from the Perspective of Embodied Intelligence: Concept, Architecture, and Value Functions},
year = {2026},
journal = {Science-Technology & Publication},
volume = {45},
number = {5},
pages = {37-48},
keywords = {embodied intelligence, generative artificial intelligence, agent, digital publishing, intelligent publishing, publishing agent, agent publishing, Outline of the 15th Five-Year Plan (2026–2030)},
url = {https://www.sciopen.com/article/10.16510/j.cnki.kjycb.20260521.005},
doi = {10.16510/j.cnki.kjycb.20260521.005},
abstract = {The Implementation Opinions on the Standardized Application and Innovative Development of Agents emphasizes strengthening application-driven development of agents, actively and prudently promoting their deployment in typical scenarios, and exploring demonstration use cases such as content creation agents, educational and instructional agents, and agents for online content construction and management. An agent is an intelligent system capable of autonomous perception, memory, decision-making, interaction, and execution. A publishing agent is such a system with domain-specific capabilities applied to the publishing field, constituting an important form of intelligent publishing products and services and a new category of product in digital publishing development. Based on this, this paper divides the architecture of a publishing agent into five modules: the publishing context perception module, the publishing task planning and reasoning module, the full-process publishing production module, the publishing knowledge memory module, and the multi-role collaboration module. Specifically, the perception module converts heterogeneous data (e.g., text, sound, images) into content understandable by large models and transmits it to subsequent reasoning and decision-making stages. The task planning and reasoning module—the agent’s cognitive core—plans, learns from, and makes decisions based on information acquired during perception. The full-process production module acts as the "operator" of the publishing agent, converting decisions from the reasoning module into concrete behaviors and actions, with key technologies including robot control, action generation, and content output. The knowledge memory module stores historical interaction data to support ongoing decision-making. The multi-role collaboration module, critical in multi-agent systems, handles interactions among users and agents, agents and agents, and agents and the environment; this paper focuses specifically on inter-agent communication. The paper further summarizes four types of publishing agents: dedicated agents for specific publishing processes, autonomous agents that learn from feedback in publishing scenarios, generative agents serving publishing production, and multi-agent systems covering the entire publishing workflow. Finally, from the perspective of "space-subject-element", the paper provides an in-depth analysis of the value and functions of publishing agents in expanding the space of integrated publishing, innovating intelligent publishing subjects, and optimizing new elements such as data and technology.}
}