Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Collaborative LLM Agents for C4 Software Architecture Design Automation

Kamil Szczepanik, Jarosław A. Chudziak

TL;DR

The paper tackles the problem of manually crafting C4 software architecture models by introducing an LLM-based multi-agent system (MAS) that simulates expert dialogues to generate Context, Container, and Component views. It presents a top-down workflow (across L1-L3) coupled with a hybrid evaluation framework that combines deterministic checks with LLM-based qualitative scoring (LLM-as-a-Judge) to assess structure, consistency, and semantics. Key contributions include the MAS_C4 architecture, a prompt-engineering strategy for collaborative and processing agents, and an empirical study comparing four LLMs across configurations, plus a discussion of limitations and future work. Findings indicate that single-agent configurations often yield higher semantic fidelity and clarity, while multi-agent setups tend to produce broader, more complex C4 models; prompts and evaluation methods critically influence output quality, and the approach offers a fast, information-rich starting point for automated SAD. Overall, the work advances automated software architecture design by integrating collaborative AI reasoning, structured artifact pipelines, and scalable evaluation for C4 modeling.

Abstract

Software architecture design is a fundamental part of creating every software system. Despite its importance, producing a C4 software architecture model, the preferred notation for such architecture, remains manual and time-consuming. We introduce an LLM-based multi-agent system that automates this task by simulating a dialogue between role-specific experts who analyze requirements and generate the Context, Container, and Component views of the C4 model. Quality is assessed with a hybrid evaluation framework: deterministic checks for structural and syntactic integrity and C4 rule consistency, plus semantic and qualitative scoring via an LLM-as-a-Judge approach. Tested on five canonical system briefs, the workflow demonstrates fast C4 model creation, sustains high compilation success, and delivers semantic fidelity. A comparison of four state-of-the-art LLMs shows different strengths relevant to architectural design. This study contributes to automated software architecture design and its evaluation methods.

Collaborative LLM Agents for C4 Software Architecture Design Automation

TL;DR

The paper tackles the problem of manually crafting C4 software architecture models by introducing an LLM-based multi-agent system (MAS) that simulates expert dialogues to generate Context, Container, and Component views. It presents a top-down workflow (across L1-L3) coupled with a hybrid evaluation framework that combines deterministic checks with LLM-based qualitative scoring (LLM-as-a-Judge) to assess structure, consistency, and semantics. Key contributions include the MAS_C4 architecture, a prompt-engineering strategy for collaborative and processing agents, and an empirical study comparing four LLMs across configurations, plus a discussion of limitations and future work. Findings indicate that single-agent configurations often yield higher semantic fidelity and clarity, while multi-agent setups tend to produce broader, more complex C4 models; prompts and evaluation methods critically influence output quality, and the approach offers a fast, information-rich starting point for automated SAD. Overall, the work advances automated software architecture design by integrating collaborative AI reasoning, structured artifact pipelines, and scalable evaluation for C4 modeling.

Abstract

Software architecture design is a fundamental part of creating every software system. Despite its importance, producing a C4 software architecture model, the preferred notation for such architecture, remains manual and time-consuming. We introduce an LLM-based multi-agent system that automates this task by simulating a dialogue between role-specific experts who analyze requirements and generate the Context, Container, and Component views of the C4 model. Quality is assessed with a hybrid evaluation framework: deterministic checks for structural and syntactic integrity and C4 rule consistency, plus semantic and qualitative scoring via an LLM-as-a-Judge approach. Tested on five canonical system briefs, the workflow demonstrates fast C4 model creation, sustains high compilation success, and delivers semantic fidelity. A comparison of four state-of-the-art LLMs shows different strengths relevant to architectural design. This study contributes to automated software architecture design and its evaluation methods.
Paper Structure (17 sections, 4 equations, 4 figures, 2 tables)

This paper contains 17 sections, 4 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Multi-Agent LLM Architecture
  5. System State $\boldsymbol{S}$
  6. Artifact Set $\boldsymbol{\mathcal{A}}$
  7. Agent Set $\boldsymbol{A}$
  8. Collaborative Analysis Agents ($A_{collab}$)
  9. Specialized Processing Agents ($A_{proc}$)
  10. C4 Model Generation Workflow
  11. Prompt Engineering
  12. Evaluation Module
  13. Experimental Setup
  14. Results
  15. Discussion
  16. ...and 2 more sections

Figures (4)

  • Figure 1: The proposed system workflow, illustrating the multi-agent C4 model generation pipeline and the subsequent evaluation framework. The generation stage (left) shows expert agents collaboratively processing a system brief, with specialized agents producing C4 artifacts and diagrams, while the evaluation stage (right) assesses these artifacts using structural, rule-based, and semantic criteria.
  • Figure 2: Top-down multi-level decomposition workflow for C4 architecture modeling.
  • Figure 3: System brief input of the Library Management System test case.
  • Figure 4: Generated artifacts for the Library Management System test case: an excerpt from the Analysis Report (a) and its corresponding PlantUML diagram (b) at L2 (Container). Artifacts were produced by the system using Grok 3 mini in the Collaborative Analysis (1 Round) configuration.