-
Background: As software systems grow in size and complexity, traditional Continuous Integration and Continuous Deployment/Delivery (CI/CD) pipelines face significant efficiency bottlenecks. Although researchers have begun using Large Language Models (LLMs) to optimize pipeline tasks, these intelligent agents are rarely integrated directly into pipeline control flows or gating decisions. This disconnection stems from a focus on agent design at the expense of understanding internal pipeline dynamics and structural constraints, limiting practical