Production Agent Patterns

The only repo that shows the same agent implemented across ALL major frameworks with production deployment configs.

🎯 What This Repo Provides

1. Reference Implementations: Same agent built across 6 frameworks for direct comparison
2. Production Configs: Real deployment code, not localhost demos
3. MCP Servers: Reusable tool integration patterns
4. Cost Analysis: Actual production cost data
5. Evaluation Suites: Standard testing patterns

📊 The 7 Pillars Framework

This repository implements the 7 Pillars of Production Agent Systems:

Pillar	What It Solves
Orchestration	Multi-agent coordination
Memory	State persistence across sessions
Guardrails	Safety and validation
Observability	Tracing and debugging
Security	Access control and audit
Cost Management	Token and resource budgets
Lifecycle (AgentOps)	CI/CD for agents

🏗️ Repository Structure

production-agent-patterns/
├── docs/                           # Framework documentation
│   ├── 7-pillars.md               # The core framework
│   ├── provider-comparison.md      # Detailed matrix
│   └── decision-guide.md          # How to choose
│
├── agents/                         # Reference agent implementations
│   └── research-assistant/        # Primary reference agent
│       ├── openai-sdk/            # OpenAI Agents SDK
│       ├── claude-sdk/            # Claude Agent SDK
│       ├── langraph/              # LangGraph
│       ├── aws-strands/           # AWS Strands
│       ├── google-adk/            # Google ADK
│       └── oracle-adk/            # Oracle ADK
│
├── mcp-servers/                    # Model Context Protocol servers
│   ├── template/                  # Starter template
│   └── database-connector/        # PostgreSQL/MySQL example
│
├── deployment/                     # Infrastructure as Code
│   ├── aws/                       # Terraform + CDK
│   ├── gcp/                       # Terraform
│   ├── azure/                     # Bicep + Terraform
│   ├── oracle/                    # OCI Terraform
│   └── docker/                    # Local development
│
├── monitoring/                     # Observability setup
│   ├── langfuse/                  # Open-source tracing
│   └── dashboards/                # Grafana dashboards
│
└── evaluation/                     # Testing and benchmarks
    ├── test-suites/               # Standard evaluation patterns
    └── benchmarks/                # Performance baselines

🚀 Quick Start

1. Clone and Setup

git clone https://github.com/frankxai/production-agent-patterns.git
cd production-agent-patterns

# Choose your framework
cd agents/research-assistant/openai-sdk  # or claude-sdk, etc.

# Install dependencies
pip install -r requirements.txt

# Set environment variables
cp .env.example .env
# Edit .env with your API keys

2. Run the Reference Agent

# OpenAI version
python main.py "Research the latest developments in quantum computing"

# Claude version
cd ../claude-sdk
python main.py "Research the latest developments in quantum computing"

3. Compare Implementations

All implementations produce the same output format, making direct comparison possible.

🔧 The Research Assistant Agent

Our reference agent is a Research Assistant that:

• Searches the web for information
• Reads and summarizes documents
• Synthesizes findings into reports
• Cites sources

This agent is complex enough to demonstrate all 7 pillars while simple enough to understand.

Agent Specification (Framework-Agnostic)

agent:
  name: ResearchAssistant
  description: Researches topics and produces synthesized reports

  tools:
    - web_search: Search the web for information
    - fetch_url: Retrieve and parse web pages
    - summarize: Condense long documents

  memory:
    short_term: conversation context
    long_term: research history, user preferences

  guardrails:
    input: content_policy, pii_detection
    output: citation_required, format_validation

  output_format:
    - summary: 2-3 paragraph synthesis
    - key_findings: bullet points
    - sources: cited URLs

📦 MCP Server Templates

Build once, use with any agent framework.

Database Connector

# mcp-servers/database-connector/server.py
from mcp import Server, Tool

server = Server("database-connector")

@server.tool()
async def query_database(sql: str) -> dict:
    """Execute read-only SQL query"""
    # Implementation with safety checks
    return execute_safe_query(sql)

🌐 Deployment

AWS (Bedrock + AgentCore)

cd deployment/aws
terraform init
terraform apply -var="agent_name=research-assistant"

Google Cloud (Vertex AI)

cd deployment/gcp
terraform init
terraform apply

Azure (AI Foundry)

cd deployment/azure
az deployment group create \
  --resource-group myResourceGroup \
  --template-file main.bicep

Oracle (OCI + ADK)

cd deployment/oracle
terraform init
terraform apply -var="compartment_id=ocid1.compartment..."

Local Development

cd deployment/docker
docker-compose up

📈 Monitoring

Langfuse Setup

cd monitoring/langfuse
docker-compose up -d

# Open http://localhost:3000
# Default credentials: admin / admin

Import Dashboards

cd monitoring/dashboards
./import-to-grafana.sh

🧪 Evaluation

Run Test Suite

cd evaluation
pytest test-suites/ -v --benchmark

Test Categories

Suite	Tests
Functional	Agent produces correct outputs
Safety	Guardrails block bad inputs/outputs
Performance	Latency and throughput
Cost	Token usage tracking
Adversarial	Red team attack resistance

💰 Cost Comparison

Based on 1,000 research queries (as of January 2026):

Framework	Avg Cost/Query	Latency (p50)	Notes
OpenAI SDK	$0.08	2.3s	GPT-4o
Claude SDK	$0.07	2.1s	Claude Sonnet 4
AWS Bedrock	$0.06	2.5s	Claude via Bedrock
Google ADK	$0.07	2.4s	Gemini 2.0
Oracle ADK	$0.05	2.6s	Cohere Command A

Costs include all API calls, tool usage, and retry logic.

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Priority Areas

1. Additional agent implementations (AutoGen, CrewAI)
2. More MCP server examples
3. Cost optimization patterns
4. Evaluation benchmark improvements

📚 Related Resources

• The 7 Pillars Blog Post
• Model Context Protocol
• Provider Documentation Links

📄 License

MIT License - see LICENSE for details.

---

Built with ❤️ by Frank | AI Architect

Helping you ship production agents, not just demos.