Intel / article

Towards a Characterization of Microservice Architectures Generated by Large Language Models

Source: arxiv.org
Author: Jos\'e Renan, Ademar Sousa, Emanuel Dantas, Danyllo Albuquerque, Mirko Perkusic, Kyller Gorg\^onio, Angelo Perkusich
Date: Published Sep 18, 2026

Context

Examples in a prompt can change how AI proposes splitting an application into smaller services. Researchers gave models from OpenAI and DeepSeek descriptions of two applications organized as modular monoliths, then compared prompts with and without examples. They measured the structure and descriptions of the resulting service designs. Study abstract.

Prompts containing examples consistently produced finer divisions, less dense communication between services and more detailed responsibility descriptions. Prompting strategy accounted for more of the observed differences than model provider. Because the study measured these structural properties rather than architectural correctness, the finding shows how examples steer a design proposal, without establishing that the resulting design is better. Findings and scope.

Key takeaways · AI-distilled

Comparing zero-shot and few-shot prompting across GPT-based and DeepSeek models, the study found architecture differences were driven mainly by prompting strategy, not which model provider was used.
Few-shot prompting consistently produced finer-grained service decomposition, lower inter-service communication density, and more detailed responsibility descriptions than zero-shot.
Zero-shot prompting instead favored coarser, more tightly coupled microservice designs.
The study measured structural properties only, not architectural correctness or quality -- it's offered as a baseline for future work on -driven architecture generation.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.

Read the source arxiv.org