TLDR: MAAD (Multi-Agent Architecture Design) is a new framework that uses a team of four specialized AI agents (Analyst, Modeler, Designer, Evaluator) to automate software architecture design. It interprets requirements, creates architectural blueprints, and provides evaluation reports. MAAD outperforms existing multi-agent systems like MetaGPT in completeness and detail. It also shows that integrating external knowledge improves design quality and that the choice of the underlying AI model significantly impacts the outcome. The research highlights the potential of AI for architecture design while emphasizing the continued need for human oversight.
Software architecture design is a crucial but often complex and time-consuming part of developing any software. It demands deep expertise, experience, and careful balancing of various quality aspects, often leading to limited design options, especially under tight deadlines. While Large Language Models (LLMs) have shown great promise in many software engineering tasks, their use in architecture design has been less explored, particularly given the need for diverse knowledge and complex decision-making.
Relying on a single LLM agent can sometimes lead to inconsistent or unreliable results, especially for tasks that require collaborative thinking. To tackle these challenges, researchers have introduced MAAD (Multi-Agent Architecture Design), an automated framework that uses a knowledge-driven Multi-Agent System (MAS) for software architecture design.
How MAAD Works: A Team of Specialized AI Agents
MAAD operates by orchestrating four specialized AI agents, each with a distinct role, to collaboratively interpret software requirements and produce detailed architectural blueprints, complete with evaluation reports based on quality attributes. These agents mimic the collaborative dynamics of human design teams, with each focusing on a specific subtask:
- Analyst Agent: This agent is responsible for understanding the software requirements, identifying key aspects, and breaking them down into functional, non-functional, architecturally significant requirements, and design constraints.
- Modeler Agent: Taking the refined requirements from the Analyst, the Modeler agent builds a multi-view representation of the system’s architecture, following the widely used “4+1” architecture view model. This includes various UML diagrams like Class, Component, and Deployment diagrams.
- Designer Agent: Based on the architectural views, the Designer agent generates detailed architecture documentation. This includes defining system goals, detailed design specifications, component and connector specifications, key technologies, design decisions, and even a skeleton of the executable prototype code.
- Evaluator Agent: The final agent rigorously assesses the generated architecture against the original requirements. It produces an ATAM (Architecture Tradeoff Analysis Method) evaluation report and a Mismatch Analysis report, highlighting any deviations or trade-offs made during the design process.
A key feature of MAAD is its knowledge-driven approach. Both the Modeler and Designer agents can access an external knowledge base, which includes authoritative literature and domain-specific best practices. This allows the agents to make informed decisions, reduce informational gaps, and ensure the generated designs align with industry standards and proven architectural patterns.
MAAD’s Performance and Advantages
The research evaluated MAAD through a case study and comparative experiments against MetaGPT, another state-of-the-art MAS. The results showed MAAD’s superiority in generating comprehensive architectural components and delivering insightful, structured architecture evaluation reports. For instance, MAAD’s Analyst agent produced a much finer granularity and greater coverage of requirements compared to MetaGPT.
In terms of system design, MAAD generated complete “4+1” architectural view models, including various detailed UML diagrams, whereas MetaGPT only produced basic class and sequence diagrams. Furthermore, MetaGPT lacked a comprehensive evaluation mechanism, a core strength of MAAD through its Evaluator agent.
Feedback from industrial architects, based on 11 real-world requirements specifications, further reinforced MAAD’s practical usability. Architects found MAAD’s generated models to align with design principles and highlighted the usefulness of its detailed mismatch reports. They particularly praised MAAD’s ability to leverage extensive external knowledge, noting that an LLM can recall knowledge much more comprehensively and rapidly than a human.
The Impact of External Knowledge and LLM Choice
The study also explored how external knowledge influences design quality. Integrating authoritative literature into MAAD’s knowledge base led to more modular and interface-driven architecture designs, aligning better with established architectural best practices. This highlights the importance of domain grounding in automated architectural design.
Additionally, the researchers investigated the performance of MAAD with three different LLMs: GPT-4o, DeepSeek-R1, and Llama 3.3. While all three produced functionally plausible architectures, GPT-4o exhibited better performance in producing architecture designs, particularly in terms of lower mismatch rates with the original requirements. This finding underscores that the choice of the underlying LLM significantly impacts the quality of the generated artifacts and that LLM selection should be task-specific.
Also Read:
- Automating Chip Verification with a Multi-Agent AI System
- AgentMesh: An AI Framework for Automated Software Creation
Future Outlook
While MAAD represents a significant step towards automating architecture design, the researchers acknowledge that it might have limitations when applied to highly complex industrial systems. Future work aims to enhance MAAD by integrating memory mechanisms and orchestrating an even more granular set of agents to better address intricate and large-scale software requirements.
For more details, you can read the full research paper here.
