Intelligent agents can autonomously perceive, reason logically, coordinate across domains, and execute tasks in a closed loop. Photo: VCG

Advances in large language models (LLMs) have accelerated the deployment of intelligent agents—advanced artificial intelligence (AI) systems capable of acting autonomously—across a wide range of government functions. Unlike traditional, passive information systems, intelligent agents can autonomously perceive, reason logically, coordinate across domains, and execute tasks in a closed loop. They have evolved from administrative support tools into autonomous actors within digital governance systems, taking on an increasing share of routine administrative tasks. This development fundamentally challenges the long-standing assumption in public administration that humans are the sole actors in governance, reshaping the field’s theoretical foundations, government organizational behavior, citizen experiences, and research methods and technologies. As AI helps modernize national governance, the growing use of intelligent agents is deepening existing lines of research in public administration and opening up new ones. Drawing on behavioral public administration (BPA), this article examines the key behavioral characteristics of intelligent agents and considers how their emergence is reshaping research agendas across major areas of public administration.

Expanding range of actors, reshaping theoretical frameworks

BPA is an interdisciplinary field at the intersection of public administration and psychology, using scientific methods to examine where public administration phenomena and psychological processes overlap. Behavior is both its starting point and its ultimate focus, with behavioral issues in public administration settings examined through the behavioral sciences, particularly psychology and behavioral economics. Central to these settings are interactions between governments and citizens, which underpin effective public governance. BPA research examines how government behavior and citizen experiences influence each other. Its theoretical framework can be summarized as a “government behavior–citizen experience” two-wheel model, which has been used to analyze a wide range of public administration issues.

Traditionally, governments and citizens have been the only two actors in government–citizen interactions. The emergence of intelligent agents, however, may fundamentally disrupt this conventional relationship. In normative terms, participants in these interactions must be capable of both judgment and action. Intelligent agents can already exercise judgment and make decisions autonomously within predefined procedural boundaries, performing routine administrative tasks—work to which AI is particularly well suited. These tasks recur frequently, follow relatively clear rules, and have a direct bearing on citizens or businesses, making the effects of AI adoption comparatively easy to observe. Agents have also moved beyond functioning solely as technological tools: They can now act autonomously without direct human control, albeit in pursuit of public objectives. As governments adopt digital and intelligent technologies, agents thus have the capabilities needed to participate as actors in public administration. Government–citizen interaction must therefore be understood through an expanded configuration—“government + intelligent agents–citizens”—that includes interactions between agents and citizens as well as integration between government personnel and agents. Accordingly, BPA’s two-wheel model should evolve into a three-gear model.

Enabling human–agent interaction, reshaping citizen experiences

A central question in citizen–agent interactions in public administration is how citizens’ cognition and emotions change when they engage with government service agents rather than human officials. Some scholars argue that AI-assisted decision-making can keep officials’ personal biases from influencing decisions, while algorithmic design can ensure greater consistency and predictability. If frontline civil servants are replaced by AI agents, will administrative discretion diminish? Will citizens accept agents’ decisions more calmly and comply more readily? Will trust develop, be maintained, and erode differently when citizens encounter decisions made by AI civil servants rather than human officials?

Making decisions also entails taking responsibility for them. Intelligent agents pose a particularly acute accountability dilemma because of their limitations as autonomous actors—they lack self-awareness and the capacity for reflection, and have yet to acquire human-like emotional understanding. Mind perception theory distinguishes between two principal psychological capacities: agency and experience. The former concerns the capacity for intention and action, whereas the latter concerns the capacity for sensation and emotion. At present, intelligent agents are considered to possess some degree of agency but relatively limited experiential capacity. These limitations raise questions about how citizens attribute responsibility and seek accountability when agents make errors or fail to deliver public services, what emotions such failures provoke, and what consequences follow. Beyond making public decisions at various levels, governments have a core responsibility to provide a wide range of public services. If service provision shifts from a predominantly government-led model to one involving both government and intelligent agents, or even becomes agent-led at certain stages, citizens’ participation in coproduction and value co-creation may change in unintended ways.

Pursuing human–agent integration, reshaping government behavior

In 2025, the emergence of the Chinese-developed generative AI (GenAI) model DeepSeek triggered a nationwide wave of intelligent upgrades to government services, marking a shift from digitalization toward the use of intelligent systems. AI applications have rapidly expanded across the public sector, while relationships between civil servants and intelligent agents have gradually evolved from human–AI collaboration toward human–AI integration. Whether this integration takes hold ultimately depends on individuals’ attitudes toward agents and how they work with them.

Research in this area increasingly examines how civil servants perceive smart technology, AI, robotics, and algorithms (STARA), and how this “STARA awareness” influences organizational behavior in the public sector. STARA awareness comprises two dimensions: hindrance appraisal, which frames STARA as a threat, and challenge appraisal, which frames it as an opportunity. Civil servants exhibit patterns of STARA awareness that differ from those of employees in private enterprises and other nonpublic organizations. The macro-level institutional, meso-level organizational, and micro-level individual factors shaping these patterns therefore warrant systematic investigation.

When hindrance appraisal dominates, STARA awareness may lead to adverse consequences, including AI anxiety, job insecurity, administrative burden, blame-avoidance behavior, and turnover intentions. By contrast, when challenge appraisal dominates, it may encourage work engagement, responsible action, innovative behavior, proactive change behavior, and organizational citizenship behavior. The mechanisms linking STARA awareness to one or more outcomes may involve double-edged effects or dual pathways. Civil servants’ patterns of work behavior may also differ between the human–AI collaboration and human–AI integration stages. The effects of errors made by intelligent agents on both collaboration and integration also deserve particular attention.

Advancing methodological innovation and reshaping research paradigms

Experimental methods are a defining feature of BPA research. Over the past decade, BPA has helped drive the wider use of experiments in public administration, strengthening the scientific rigor and methodological consistency of research across the discipline. Nevertheless, experimental research in public administration remains dominated by survey experiments, particularly vignette experiments. Laboratory experiments, which generally offer higher internal validity, and field experiments, which generally offer higher external validity, remain relatively scarce, leaving considerable scope for their wider use.

Experimental methods depend heavily on scenario design and experimental control. Using GenAI in experiments therefore requires both well-designed scenarios and rigorous control. LLM-based GenAI can comprehensively recreate the scenarios used in public administration experiments, including those presented through text, images, audio, video, and augmented or virtual reality. This can make scenarios feel more realistic to participants and strengthen external validity. With appropriate prompts, researchers can also quickly generate experimental materials tailored to specific requirements. GenAI can help keep these materials identical except for variations in the independent variable, thereby safeguarding internal validity.

The appropriate use of “silicon samples” in social science research has also drawn considerable attention and sparked academic debate. Silicon samples are virtual respondents generated by LLMs to simulate human attitudes and behaviors in surveys or behavioral experiments. Under idealized conditions, researchers could provide LLMs with the characteristics of human respondents to generate participants precisely matched to a study’s requirements, potentially replacing human samples. Silicon samples offer clear advantages over human samples in scale and speed. However, because social science investigates human society, researchers should use them cautiously. Silicon samples can be valuable in pilot studies and replication experiments, but formal studies should continue to rely primarily on human participants. Finally, both generating experimental materials and using silicon samples can cost less than traditional approaches, adding to GenAI’s appeal for researchers.

Intelligent agents continue to reshape public administration as a discipline, with implications for BPA that reach beyond the four areas examined here—systematically revising theoretical models, broadening the scope of research on citizen experiences, reconstructing the mechanisms underlying government behavior, and developing new research methods. At a more fundamental level, public administration in the AI era moves beyond the traditional study of human–human and human–organization relationships toward a new analytical framework of human–agent–organization interactions. The field is thus moving from approaching AI to engaging directly with it.

Future BPA research should be grounded in the realities of governance in China and remain committed to explaining behavioral phenomena in public administration settings. Researchers should integrate established methods with cutting-edge technologies, using AI as a research tool while also studying intelligent agents themselves and their influence on behavior in public administration. Such work can help develop public administration theories with Chinese characteristics, contribute to fostering China’s independent knowledge system in philosophy and social sciences, and provide practical guidance for modernizing national governance.

Zhang Shuwei is a professor from the School of Government at Sun Yat-sen University.