AI Agents as Infrastructures: Mediating Power Dynamics in Global AI Platform Ecosystems between the US and China

Previous research has shown that the political economy of Generative AI (GenAI) is shaped by infrastructural dominance of US-based Big Tech companies—Google, Microsoft, Amazon—which control cloud computing, data access, proprietary software, and pools of AI talent. However, the emergence of task-oriented AI agents and the growing influence of Chinese tech firms potentially reshuffle this political economy. AI agent systems, such as GPT Builder (OpenAI), Copilot Studio (Microsoft), and Coze (ByteDance), integrate foundation models with complementary components to perform specific tasks in programmable, modular, and increasingly customized ways. As such, AI agents are not only reconfiguring the technical architectures of GenAI development pipelines, but are also operating as infrastructural middleware towards the platformization and customizable industrialization of GenAI.

These shifts call for a renewed examination of infrastructural mediation and control of GenAI platform ecosystems across both U.S. and Chinese contexts. Hence, this study aims to investigate how AI agents are mediating technical interdependencies and power asymmetries within the global political economy of GenAI by asking: What are the technical architectures of AI agent systems in the U.S. and China? How do these systems operate differently as infrastructural components within broader AI ecosystems? And how do they reflect or reshape the shifting political-economic relations?
To address these questions, this study will conduct a comparative case study of two leading AI agent builders—GPT Builder from the U.S. and Coze from China. Employing a “technographic” approach, it will critically scrutinize publicly available materials including products pages, blog posts, press releases and relevant media and industry reports. The analysis will first map distinct orchestrations of core components such as foundation models, tools, memory systems, knowledge base and prompting on each builder. Based on this computational overview, this study will then situate these building systems within their respective platform architectures and AI technology stacks in each country to unfold the underlying infrastructural relations and power structures.

By examining the institutional distinctions between two cases, this study will offer empirical insight into how OpenAI and ByteDance are institutionalizing distinct modes of infrastructural dominance and interdependence. It argues that the political economy of GenAI is being reconfigured through an increasingly contested geopolitical terrain, where AI agents function as a new infrastructural layer enabling ecosystem-based digital dependency and rentier economy. These new dynamics extend infrastructural control beyond the lock-ins in cloud infrastructures of US-based tech companies, signalling a shift toward customized industrialization and governance over downstream ecosystems. In doing so, this research will lay a foundation for further critical inquiry, policy regulation and public engagement on AI governance and platform power.