From Feature Variability to Agency Variability: A Software Product-Line Engineering Framework for Governed Agentic AI Systems

Agentic artificial intelligence systems increasingly combine language models, memory, retrieval, tool use, orchestration, and human oversight. For software engineering, this creates a variability problem that feature-oriented product line methods only partly address: organizations are configuring not only functions or components, but permitted patterns of agency. Unlike MAS-SPL, which mainly structures families of agent roles and interactions, the proposed approach targets LLM-based agents whose prompts, retrieval sources, tool authority, runtime monitoring, and governance boundaries vary together. This paper proposes Agency-Centric Product Line Engineering for governed agentic AI systems. It defines agency variability as systematic variation in a system’s capacity to pursue goals, perceive context, reason and plan, use tools, exercise authority, interact with humans or other agents, remain observable, and evolve under governance constraints. It also defines semantic observability as runtime monitoring of whether semantic behavior—goals, tool choices, evidence use, proposed actions, and escalation decisions—remains consistent with the granted agency profile. The paper contributes a nine-dimension Agency Variability Model, an Agency Configuration Schema, a constraint-based validation algorithm, and a prototype-style derivation of portfolio-management agent variants. The case shows that five variants can share assets while differing in perception, authority, autonomy, human control, and topology.