Deep Expertise Track · Lesson 12
Capstone Project
Combine everything into one multi-agent BA system
Capstone: Build a Multi-Agent BA System
Lesson 12 — combine everything you've learned into one project
What you'll learn
- How to choose the right orchestration pattern for a real problem
- Combine sequential + concurrent + handoff patterns in one system
- Build a production-quality multi-agent BA system you can show in interviews
- Write a LinkedIn post about what you built and learned
The Problem
Build a multi-agent system that does end-to-end BA work: given a feature request, it:
- Triage the request (handoff pattern) → route to the right specialist
- Analyze from multiple perspectives (concurrent pattern) → requirements + process + risk
- Synthesize into a BRD (sequential pattern) → draft → review → polish
- Evaluate the output (LLM-as-judge) → score quality
┌──────────────────────────────────────────────────────────────┐
│ CAPSTONE: BA MULTI-AGENT SYSTEM │
│ │
│ Feature Request: "Add OTP login to mobile app" │
│ │ │
│ ▼ │
│ ┌──────────────┐ │
│ │ TRIAGE AGENT │ (HANDOFF pattern) │
│ │ "Routes to │── "This needs requirements + risk analysis"│
│ │ specialists" │ │
│ └──────┬───────┘ │
│ │ │
│ ▼ (CONCURRENT pattern — fan-out) │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │Requirements │ │Process Map │ │Risk Assess │ │
│ │ Agent │ │Agent │ │ Agent │ │
│ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │
│ │ │ │ │
│ └────────────────┼─────────────────┘ │
│ │ (fan-in) │
│ ▼ │
│ ┌──────────────────────────────────────────┐ │
│ │ SYNTHESIS PIPELINE (SEQUENTIAL pattern) │ │
│ │ │ │
│ │ Draft Agent ──▶ Review Agent ──▶ Polish │ │
│ │ Agent │ │
│ └──────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌──────────────┐ │
│ │ EVALUATOR │ (LLM-AS-JUDGE) │
│ │ Scores 1-10 │ │
│ └──────────────┘ │
│ │ │
│ ▼ │
│ Final BRD + Quality Score │
└──────────────────────────────────────────────────────────────┘
What You've Built Across 12 Lessons
| Lesson | Concept | Pattern |
|---|---|---|
| 1 | Agent spectrum (5 levels) | Foundation |
| 2 | ReAct loop from scratch | Single agent |
| 3 | LangChain abstractions | Single agent (framework) |
| 4 | Tool design (ACI) | Cross-cutting |
| 5 | Sequential orchestration | Multi-agent: pipeline |
| 6 | Concurrent orchestration | Multi-agent: parallel |
| 7 | Handoff orchestration | Multi-agent: dynamic |
| 8 | Group chat (maker-checker) | Multi-agent: debate |
| 9 | Magentic orchestration | Multi-agent: adaptive |
| 10 | Agent evaluation | Production: quality |
| 11 | Production considerations | Production: ops |
| 12 | Capstone (this lesson) | ALL combined |
Build Steps
- Create the project:
mkdir -p ~/Documents/Pythonproject/Touseef_Project_Work/ba-multi-agent - Triage agent: Takes feature request, outputs specialist routing (Lesson 7)
- 3 specialist agents: Requirements, Process, Risk — run concurrently (Lesson 6)
- Synthesis pipeline: Draft → Review → Polish (Lesson 5)
- Evaluator: LLM-as-judge scores the final BRD (Lesson 10)
- Add cost tracking: Log token usage per agent (Lesson 11)
- Add context summarization: Between concurrent results and synthesis (Lesson 11)
- Test it: Run with 5 different feature requests. Check traces for failure modes (Lesson 10)
LinkedIn Post Template
Post after completing the capstone
Title: "I built a multi-agent AI system that does end-to-end BA work. Here's what I learned."
Body structure:
- What I built: triage → 3 concurrent specialists → sequential synthesis → LLM evaluation
- The 5 orchestration patterns I learned (name them — hiring managers search for these terms)
- The biggest challenge: context management (lessons 10-11)
- Before/after comparison: my BA Assistant (Level 1 chatbot) vs this multi-agent system (Level 5)
- What I'd improve next: LangGraph for stateful orchestration, LangSmith for tracing
- GitHub link to the repo
What to Add to Your Resume After This
RESUME ADDITIONS (after completing all 12 lessons):
SKILLS SECTION:
Agentic AI, AI Agent Orchestration, Multi-Agent Systems
ReAct Pattern, LangChain, LangGraph
Sequential, Concurrent, Handoff, Group Chat, Magentic patterns
LLM Evaluation (LLM-as-Judge), Hallucination Detection
AI Governance (from your daily practice plan)
EXPERIENCE SECTION:
"Built multi-agent BA system using LangChain with 5
orchestration patterns: triage (handoff), parallel analysis
(concurrent), synthesis pipeline (sequential), quality
evaluation (LLM-as-judge). DeepSeek API, 6 tools,
context management between agents."
PORTFOLIO:
GitHub: ba-multi-agent (3rd live project)
LinkedIn post documenting the architecture
The one-sentence summary
You now have everything you need: 5 orchestration patterns, tool design principles, evaluation methods, and production concerns — combine them into one capstone project that proves you can architect and build multi-agent AI systems.
Final Checklist
| Item | Status |
|---|---|
| Understand 5 levels of AI systems | ☐ Done |
| Built ReAct agent from scratch | ☐ Done |
| Built LangChain agent with tools | ☐ Done |
| Improved tool descriptions (ACI) | ☐ Done |
| Built sequential pipeline | ☐ Done |
| Built concurrent (parallel) system | ☐ Done |
| Built handoff (triage) system | ☐ Done |
| Built maker-checker loop | ☐ Done |
| Understand magentic pattern | ☐ Done |
| Evaluated agent with LLM-as-judge | ☐ Done |
| Understand production concerns | ☐ Done |
| Capstone multi-agent system | ☐ Done |
| LinkedIn post published | ☐ Done |
| Resume updated with agentic AI skills | ☐ Done |
Want to see these patterns in action?
Explore the live apps built with these agent architectures.
Explore the Lab →