ClickShop: Agentic Architecture Evolution Demo

A production-oriented demonstration of scaling agentic systems from MVP to 50K orders/day

Overview • Architecture • Quick Start • Demos • Resources

Overview

ClickShop demonstrates the architectural evolution of agentic systems through three production-grade implementations. Built for AWS re:Invent 2025 (DAT309 Chalk Talk), this project showcases scaling patterns from a weekend MVP to enterprise-scale multi-agent orchestration with semantic search capabilities.

⚠️ Educational Use Only: This demonstration is designed for learning purposes and illustrates production patterns without production-level error handling, monitoring, or security hardening.

The ClickShop Story

A live-streaming shopping platform that evolved from 50 orders/day to 50,000 orders/day through three architectural iterations—demonstrating how thoughtful design enables exponential scaling without complete rewrites.

Architecture Evolution

Performance & Scale Comparison

Metric	Phase 1: Single Agent	Phase 2: Agent + MCP	Phase 3: Multi-Agent
Daily Capacity	50 orders	5,000 orders	50,000 orders
Response Time	~2.0s	~3.5s	~200ms
Architecture	Monolithic	MCP-mediated	Supervisor pattern
Database Access	Direct (`psycopg3`)	RDS Data API (MCP)	Direct (`psycopg3`) + pgvector
Search Type	Exact match	Exact match	Semantic (vector)
Coupling	Tight	Loose (MCP)	Loose + Specialized

Technical Architecture Details

Phase 1: Single Agent Architecture

Capacity: 50 orders/day | Response Time: ~2.0s

Architecture:

1 Agent: Monolithic agent handles all responsibilities
4 Custom Tools:
- identify_product_from_stream() - Product discovery
- check_product_inventory() - Real-time inventory verification
- calculate_order_total() - Price calculation with tax/shipping
- process_customer_order() - Order persistence and workflow
Database: Direct Aurora PostgreSQL access via psycopg3
Pattern: Tight coupling with inline database operations

Key Characteristics:

ReAct pattern with Chain of Thought reasoning
Sequential tool execution
Manual connection pooling
Ideal for MVPs and prototypes

Phase 2: MCP-Powered Architecture

Capacity: 5,000 orders/day | Response Time: ~3.5s

Architecture:

1 Agent: Single agent with MCP integration
MCP Tools: Auto-discovered from awslabs.postgres-mcp-server
- query - SQL queries via RDS Data API
1 Custom Tool:
- create_order() - Order processing logic
Database: Aurora PostgreSQL accessed via MCP (stdio transport)
Pattern: Loose coupling through Model Context Protocol

Key Improvements:

Database abstraction layer via standardized MCP protocol
RDS Data API for serverless scaling
IAM-based authentication
Horizontal scaling capability
Connection pooling handled by MCP server

Implementation:

from mcp import stdio_client, StdioServerParameters
from strands.tools.mcp import MCPClient

mcp_client = MCPClient(lambda: stdio_client(
    StdioServerParameters(
        command="uvx",
        args=[
            "awslabs.postgres-mcp-server@latest",
            "--resource_arn", "arn:aws:rds:region:account:cluster:cluster-id",
            "--secret_arn", "arn:aws:secretsmanager:region:account:secret:secret-name",
            "--database", "postgres",
            "--region", "us-west-2",
            "--readonly", "True"
        ]
    )
))

Phase 3: Multi-Agent Supervisor Architecture

Capacity: 50,000 orders/day | Response Time: ~200ms

Architecture:

4 Specialized Agents:
1. Supervisor Agent 🎯 - Workflow orchestration (no tools)
2. Search Agent 🔍 - Semantic product discovery
  - Tool: semantic_product_search() (pgvector-powered)
3. Product Agent 📋 - Product information and inventory
  - Tools: get_product_details(), check_inventory_status()
4. Order Agent 🛒 - Order processing and confirmation
  - Tool: simulate_order_placement()
Database: Aurora PostgreSQL + pgvector extension
Search: Vector embeddings with HNSW index (cosine similarity)
Pattern: Supervisor orchestration with specialized agents

Key Improvements:

Parallel agent execution for sub-200ms response times
Semantic search using vector embeddings
Single Responsibility Principle per agent
Natural language product discovery
Horizontal and vertical scaling capabilities

Semantic Search Implementation:

# Vector embedding generation and similarity search
CREATE INDEX ON products USING hnsw (embedding vector_cosine_ops);

# Natural language search: "comfortable running shoes"
# Matches semantically similar products, not just keywords

Quick Start

Prerequisites

Required:

Python 3.11+
AWS Account with Amazon Bedrock access
Claude Sonnet 4.5 model access in Bedrock
AWS CLI configured with credentials

Recommended:

Familiarity with agentic frameworks (Strands, LangChain)
Understanding of Model Context Protocol (MCP)
Basic PostgreSQL and vector search knowledge

Installation

# Clone repository
git clone <REPOSITORY_URL>
cd sample-dat309-agentic-workflows-aurora-mcp/clickshop-demo

# Run automated setup
./scripts/setup.sh

# Activate virtual environment
source venv/bin/activate  # macOS/Linux
# venv\Scripts\activate    # Windows

# Configure AWS credentials
cp .env.example .env
# Edit .env with your AWS credentials and region

# Verify installation
python -m demos.run_demo

Environment Configuration

# .env file structure
AWS_ACCESS_KEY_ID=your_access_key
AWS_SECRET_ACCESS_KEY=your_secret_key
AWS_REGION=us-west-2
BEDROCK_MODEL_ID=global.anthropic.claude-sonnet-4-5-20250929-v1:0

Running the Demos

Interactive Demo Launcher

python -m demos.run_demo

Select from the interactive menu to run any architecture demonstration.

Individual Demo Execution

Phase 1: Single Agent (Monolithic)

python -m demos.phase_1_single_agent

Demonstrates:

Basic agentic loop with ReAct pattern
Chain of Thought (CoT) reasoning
Sequential tool execution
Direct database access patterns
Manual connection management

Phase 2: Agent + MCP (Scaling)

python -m demos.phase_2_agent_mcp

Demonstrates:

MCP server integration for database abstraction
RDS Data API for serverless compute
IAM-based authentication
Tool auto-discovery from MCP servers
Read-only database operations via MCP

Phase 3: Multi-Agent System (Production)

python -m demos.phase_3_multi_agent

Demonstrates:

Supervisor pattern for agent orchestration
Specialized agents with single responsibilities
Semantic search with pgvector (0.8.0+)
Parallel agent execution
HNSW vector indexing for similarity search

Project Structure

clickshop-demo/
├── requirements.txt               # Python dependencies
├── .env.example                   # Environment variable template
│
├── demos/                         # Demo implementations
│   ├── run_demo.py               # Main entry point with interactive menu
│   ├── phase_1_single_agent.py   # Monolithic agent (50 orders/day)
│   ├── phase_2_agent_mcp.py      # MCP integration (5K orders/day)
│   └── phase_3_multi_agent.py    # Multi-agent (50K orders/day)
│
├── lib/                           # Core library modules
│   └── aurora_db.py              # Database operations and utilities
│
├── scripts/                       # Automation scripts
│   └── setup.sh                  # Environment setup
│
└── data/                          # Static data
    └── products.json             # Product catalog

Technical Stack

Component	Technology	Purpose
Agent Framework	Strands	Agent orchestration, tool management, execution flow
LLM Runtime	Amazon Bedrock	Claude Sonnet 4.5 model hosting and inference
Database	Amazon Aurora PostgreSQL	Transactional data storage with serverless v2
Vector Search	pgvector 0.8.0+	Semantic search with HNSW indexing
Protocol	Model Context Protocol	Standardized tool/resource integration
MCP Server	awslabs.postgres-mcp-server	PostgreSQL access via RDS Data API
SDK	boto3	AWS service integration
Language	Python 3.11+	Implementation language

Troubleshooting

AWS Authentication Issues

# Verify AWS credentials
aws sts get-caller-identity

# Verify Bedrock access
aws bedrock list-foundation-models --region us-west-2

# Test Bedrock model invocation
aws bedrock-runtime invoke-model \
    --model-id global.anthropic.claude-sonnet-4-5-20250929-v1:0 \
    --body '{"anthropic_version":"bedrock-2023-05-31","max_tokens":100,"messages":[{"role":"user","content":"Hello"}]}' \
    --region us-west-2 \
    output.json

Python Environment Issues

# Recreate virtual environment
rm -rf venv
python3.11 -m venv venv
source venv/bin/activate
pip install -r requirements.txt --upgrade

MCP Server Connection Issues

# Verify MCP server installation
uvx awslabs.postgres-mcp-server@latest --version

# Test MCP server connectivity
uvx awslabs.postgres-mcp-server@latest \
    --resource_arn "arn:aws:rds:region:account:cluster:cluster-id" \
    --secret_arn "arn:aws:secretsmanager:region:account:secret:secret-name" \
    --database "postgres" \
    --region "us-west-2" \
    --readonly "True"

Technical Resources

Documentation & Specifications

Resource	Description
Strands Framework	Agent framework documentation and API reference
Amazon Bedrock	AWS managed LLM service documentation
Model Context Protocol	MCP specification and implementation guides
pgvector Extension	PostgreSQL vector similarity search
Aurora PostgreSQL	Serverless database documentation

AWS Blogs & Technical Articles

Supercharging Vector Search with pgvector 0.8.0 on Aurora PostgreSQL
Deep dive into pgvector 0.8.0 improvements including HNSW indexing, scalar quantization, and performance optimizations for semantic search workloads.

Supercharging AWS Database Development with MCP Servers
Guide to using Model Context Protocol servers for building AI agent applications with standardized database access patterns.

Key Learnings & Best Practices

Architectural Principles

Start Monolithic, Evolve Deliberately
Begin with a single agent to validate product-market fit. Introduce complexity only when metrics demonstrate the need.
Abstract Early, Scale Later
Introduce MCP for database abstraction before scaling bottlenecks emerge. Protocol-driven design enables horizontal scaling.
Specialize Agents by Responsibility
Multi-agent systems should follow Single Responsibility Principle. Each agent should have one clear purpose with dedicated tools.
Use Supervisor Pattern for Orchestration
Complex workflows require a supervisor agent that delegates to specialists rather than monolithic agents handling everything.

Performance Optimization

Semantic Search: pgvector with HNSW indexing dramatically improves search relevance and reduces latency
Parallel Execution: Supervisor pattern enables concurrent agent operations for sub-second response times
Connection Pooling: MCP servers handle connection lifecycle, eliminating manual pool management
IAM Authentication: Credential management via AWS IAM reduces operational overhead

Security Considerations

Use IAM roles and policies for database access (no hardcoded credentials)
Implement least-privilege access patterns for each agent/tool
Enable audit logging for all agent actions
Use AWS Secrets Manager for credential rotation

Dependencies

# Core Framework
python>=3.11
strands-framework>=1.0.0
boto3>=1.34.0

# Database
psycopg3>=3.1.0
psycopg3-binary>=3.1.0

# Vector Search (Phase 3)
sentence-transformers>=2.2.0
pgvector>=0.8.0

# MCP Protocol (Phase 2, 3)
mcp>=0.1.0

# Utilities
python-dotenv>=1.0.0
rich>=13.7.0

AWS re:Invent 2025 | DAT309 Chalk Talk

Built with production-grade patterns for educational purposes

Licensed under MIT-0 (MIT No Attribution)

Name		Name	Last commit message	Last commit date
Latest commit History 55 Commits
clickshop-demo		clickshop-demo
.gitignore		.gitignore
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md

License

aws-samples/sample-dat309-agentic-workflows-aurora-mcp

Folders and files

Latest commit

History

Repository files navigation

ClickShop: Agentic Architecture Evolution Demo

Overview

The ClickShop Story

Architecture Evolution

Performance & Scale Comparison

Technical Architecture Details

Quick Start

Prerequisites

Installation

Environment Configuration

Running the Demos

Interactive Demo Launcher

Individual Demo Execution

Phase 1: Single Agent (Monolithic)

Phase 2: Agent + MCP (Scaling)

Phase 3: Multi-Agent System (Production)

Project Structure

Technical Stack

Troubleshooting

AWS Authentication Issues

Python Environment Issues

MCP Server Connection Issues

Technical Resources

Documentation & Specifications

AWS Blogs & Technical Articles

Key Learnings & Best Practices

Architectural Principles

Performance Optimization

Security Considerations

Dependencies

About

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors 2

Languages

Packages