UC San Diego

This dissertation investigates how large-language-model-based generative agents can support autonomous network management in complex wireless systems. Modern wireless networks are increasingly heterogeneous, programmable, and difficult to manage manually. As network components, control applications, and configuration parameters interact in increasingly complex ways, future network management requires systems that can adapt to changing environments, coordinate decisions across independently developed components, and generate executable control logic from high-level objectives.This dissertation advances the thesis that autonomous network management can berealized by decomposing complex wireless networking tasks into domain-grounded generative-agent workflows that reason over demonstrations, environments, user intents, and specifications to produce adaptive strategies, coordinated decisions, and executable network logic. The dissertation develops this idea across network management and protocol design. First, it studies how generative agents can design adaptive and explainable communication protocols from limited demonstrations. Second, it studies how generative agents can synthesize deployable Open RAN xApps from high-level user requirements. Third, it studies how generative agents can detect and mitigate conflicts among independently developed xApps under dynamic network environments. Finally, it studies how generative agents can reason over protocol specifications to identify implicit configuration constraints and prevent misconfigurations.Together, these studies show that generative agents can provide a reasoning and synthesis layer for autonomous network management, reducing manual effort while improving adaptability, coordination, and generation in next-generation wireless networks.