Analysis of architecture options for foundation-model-based agents: A taxonomy and decision model

The rapid advancement of foundation models has led to a new generation of AI-based agents. While mature frameworks exist, the system-level architectural decision-making is still often ad-hoc, as architects lack a systematic, evidence-based framework for navigating the complex landscape of architectural options and their resulting trade-offs. To address this critical gap, we conducted a Systematic Literature Review (SLR) of 66 primary studies to develop two key artifacts: (1) a comprehensive taxonomy that systematically classifies architectural design options for foundation-model-based agents across functional capabilities and non-functional qualities, and (2) an accompanying decision model, presented as a matrix, that maps these options to their impact on system qualities, making critical design trade-offs explicit. The taxonomy and decision model were then validated through a multi-faceted process, including a formal survey with 12 experts from both academia and industry. This work provides architects with a common vocabulary and a structured guide for designing and reasoning about agent systems, thereby providing a rigorous foundation to address the current fragmentation in the field.