Building a Multi-Agent AI System: Lessons from LUCID

The Challenge

When our team signed up for the internal hackathon, we wanted to tackle something ambitious: could we build a system where multiple AI agents collaborate to solve complex engineering problems? The answer became LUCID.

Why Multi-Agent?

A single large language model is powerful but limited. It cannot simultaneously search code, query documentation, and reason about architecture. By splitting responsibilities across five specialised agents, each could focus on what it does best while sharing context through a unified orchestration layer.

Architecture Highlights

RAG Pipeline

We chose ChromaDB as our vector store and SentenceTransformers for embeddings. The key insight was that chunking strategy matters more than the embedding model. We experimented with overlapping chunks of 512 tokens and found this sweet spot between context preservation and retrieval precision.

Knowledge Graph

Beyond vector similarity, some questions need relationship traversal. We built a knowledge graph using NetworkX to capture how code components, documentation, and team expertise connect. When an agent needs to understand why a module exists, it can trace dependency chains rather than relying on semantic similarity alone.

Agent Orchestration

LangGraph gave us the framework to wire agents into a directed graph with conditional routing. Each agent has a clear interface: it receives context, performs its specialised task, and returns structured output that other agents can consume.

The Five Agents

1. Code Analyst — Deep code understanding and explanation
2. Documentation Writer — Automated technical documentation
3. Architecture Advisor — System design recommendations
4. Bug Detective — Intelligent bug analysis and root cause detection
5. Knowledge Navigator — Cross-domain knowledge synthesis

Key Takeaways

Agent specialisation outperforms generalism. A focused agent with a narrow prompt and relevant context produces significantly better results than one agent trying to do everything.

RAG quality is a chunking problem. How you split documents matters more than which embedding model you choose. Overlapping chunks with metadata preservation was our winning strategy.

Knowledge graphs complement vector search. Not every question can be answered with nearest-neighbour lookup. Relationship traversal fills the gap that pure semantic search leaves behind.

Tech Stack

• Backend: FastAPI for high-performance async API endpoints
• AI Framework: LangChain + LangGraph for agent orchestration
• Vector Store: ChromaDB with SentenceTransformers embeddings
• Knowledge Graph: NetworkX for relationship mapping
• Frontend: Streamlit for rapid dashboard prototyping

Looking Forward

LUCID demonstrated that multi-agent AI systems can meaningfully accelerate engineering workflows. The pattern of specialised agents collaborating through shared context is applicable well beyond hackathons. I am now exploring how these patterns can be applied at enterprise scale for developer productivity and knowledge management.