Abstract:
Modern software systems are increasingly
built as cloud-native, distributed applications, and an
increasing share of their source code is now produced with
the assistance of large language model (LLM) based coding
tools. Both shifts change the way architectural decisions
propagate into software quality. To investigate this
relationship, we conducted a mixed-methods study
combining controlled experimentation, empirical analysis,
case studies and architecture visualization. We analyzed a
sample of twenty-four software systems drawn from several
domains, covering layered monolithic, modular monolithic,
microservice and serverless event-driven architectures. In
the experimentation phase we re-implemented a common
reference application under four architectural designs and
measured the quality of the resulting systems using
ISO/IEC 25010 aligned metrics for maintainability,
reliability and performance efficiency, together with
delivery performance indicators. The empirical phase
collected static-analysis metrics, defect records and
deployment telemetry from the sampled systems and
analyzed the relationship between architecture and quality
using correlation and regression techniques. Case studies
were used to understand how architectural decisions
behave in specific real-world contexts, including one system
in which a significant proportion of new code was generated
by an AI coding assistant. Finally, architecture
visualization and distributed tracing were used to expose
dependency structures, architectural drift and latent hot
spots. Our findings suggest that architecture remains the
dominant controllable factor in software quality: designs
that enforce explicit module boundaries and low coupling
achieved higher maintainability and reliability irrespective
of whether the system was deployed as a monolith or as
services. We also found that AI-assisted development
accelerates local code production but increases
architectural conformance violations when boundaries are
not encoded in tooling. The study contributes evidence-
based guidance for architects, developers and engineering
managers working on cloud-native, AI-assisted software
systems.
