Bridging Requirements and Architecture: Multi-Agent Orchestration with External Knowledge and Hierarchical Memory
This paper introduces MAAD, a multi-agent framework that leverages retrieval-augmented generation and hierarchical memory to autonomously orchestrate specialized agents in transforming software requirements into comprehensive, quality-assessed architectural designs, demonstrating superior completeness and modularity compared to existing baselines.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Problem: Building a Skyscraper Without a Blueprint
Imagine you are asked to build a massive, complex skyscraper. You have a list of wishes from the client (the "requirements"), like "it needs to be earthquake-proof," "it must have 500 apartments," and "the lobby needs to be grand."
In the real world, a human Architect takes these wishes and draws detailed blueprints. They decide where the elevators go, how the steel beams connect, and how to balance safety with cost. This is incredibly hard work. It requires deep knowledge, and if the architect gets tired or misses a detail, the whole building could be unstable.
Recently, we gave this job to AI (specifically Large Language Models). But early attempts were like hiring a single AI that tried to do everything at once. It often got confused, forgot the client's wishes, or drew blueprints that looked okay on paper but wouldn't actually hold up a building. It was like a single person trying to be the structural engineer, the interior designer, and the safety inspector all at once, often making mistakes.
The Solution: MAAD (The AI Construction Crew)
The authors of this paper created MAAD (Multi-Agent Architecture Design). Instead of one AI doing everything, they built a team of four specialized AI robots, each with a specific job, working together like a human construction crew.
Here is how the team works:
The Analyst (The Translator):
- Job: Reads the client's messy wish list and turns it into a clear, organized checklist.
- Analogy: Imagine a translator who listens to a client say, "I want a house that feels like a castle but runs on solar power," and writes down a precise list: "1. Stone walls, 2. Solar panels on roof, 3. No basement." The Analyst makes sure nothing is missed.
The Modeler (The Draftsman):
- Job: Takes the checklist and draws the different views of the building (the floor plan, the electrical wiring, the plumbing).
- Analogy: This is the person who draws the "4+1" views. They don't just draw one picture; they draw the view for the electrician, the view for the plumber, and the view for the people living inside, making sure they all match up.
The Designer (The Builder's Guide):
- Job: Writes the detailed instruction manual for the construction crew.
- Analogy: This agent writes the "recipe" for the building. It says exactly what materials to use, how to connect the pipes, and how to secure the windows. It makes sure the blueprints can actually be built.
The Evaluator (The Safety Inspector):
- Job: This is the most important new addition. The Evaluator constantly checks the work of the other three.
- Analogy: Imagine a strict safety inspector who walks through the site at every stage. If the Analyst missed a requirement, the Evaluator says, "Hey, you forgot the fire escape!" If the Modeler drew a wall where a door should be, the Evaluator says, "Fix that." The Evaluator sends the work back to be fixed until it passes all safety checks.
The Secret Weapons: The Library and The Memory
To make sure this AI team doesn't make up fake facts (a problem called "hallucination"), MAAD uses two special tools:
The Library (RAG - Retrieval-Augmented Generation):
- How it works: Before the team draws anything, they open a digital library of famous, proven building codes and architectural rules (like ISO standards).
- Analogy: Instead of guessing how to build a bridge, the team looks up the "Best Practices for Bridges" book. This ensures they follow real-world rules rather than making things up.
The Memory System:
- How it works: The team remembers what they learned in previous projects.
- Analogy:
- Working Memory: Like a whiteboard in the meeting room where they write down what they are doing right now.
- Episodic Memory: Like a diary of past projects. "Last time we built a school, we forgot the gym. Let's not forget that this time."
- Semantic Memory: Like a permanent encyclopedia of architectural wisdom that the whole team can read to learn general rules.
What They Found (The Results)
The researchers tested MAAD against a popular AI tool called MetaGPT (which is like a general-purpose AI team that tries to write code and design software but isn't specialized for architecture).
- Better Blueprints: MAAD produced much more complete and organized blueprints. While MetaGPT often forgot to define how parts of the building connect (like missing doorways), MAAD defined every connection clearly.
- Fewer Mistakes: Because of the Evaluator agent, MAAD caught its own mistakes and fixed them before finishing. MetaGPT often just handed over a blueprint with errors.
- The "Brain" Matters: The researchers tested MAAD with different "brains" (different AI models). They found that the team only worked well if the "brain" was smart enough.
- Smart Brains (GPT-5.2, Qwen3.5): These models understood the complex rules and created detailed, professional-grade blueprints.
- Less Smart Brains (DeepSeek-R1, Llama3.3): These models created very simple, almost toy-like blueprints. They missed the complex details, even though they had access to the same library.
The Bottom Line
The paper concludes that MAAD is a powerful tool, but it's not a magic wand that replaces human architects entirely.
- What it does well: It can take a rough idea and turn it into a structured, detailed, and checked set of blueprints much faster than a human starting from scratch. It handles the heavy lifting of organizing information and checking for consistency.
- What it needs: It still needs a human architect to look over the final work. The AI is great at following rules and organizing data, but humans are still needed to make the final judgment calls, especially for tricky situations where "gut feeling" or hidden experience matters.
In short, MAAD is like a highly skilled, super-organized assistant architect who never gets tired, always checks the rules, and remembers every past project, but still needs a human boss to give the final approval.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.