Complete Agent Example¶

"Putting it all together: A production-inspired agentic system"

📖 Overview¶

This example demonstrates how to combine multiple agentic patterns into a cohesive, production-ready system inspired by Codex's architecture.

🎯 Patterns Integrated¶

This complete example integrates:

1. ✅ Prompt Chaining - Multi-turn conversation
2. ✅ Routing - Dynamic tool dispatch
3. ✅ Parallelization - Concurrent tool execution
4. ✅ Tool Use - External system integration
5. ✅ Memory Management - Conversation persistence
6. ✅ Exception Handling - Retry logic and recovery
7. ✅ Human-in-the-Loop - Approval workflows
8. ✅ Guardrails - Safety checks and sandboxing

🏗️ Architecture¶

┌─────────────────────────────────────────────────────────────┐
│                    Codex-Inspired Agent                      │
├─────────────────────────────────────────────────────────────┤
│                                                               │
│  ┌───────────────────────────────────────────────────────┐  │
│  │            Conversation Manager                        │  │
│  │  • Turn-based execution loop                          │  │
│  │  • History management                                 │  │
│  │  • State persistence                                  │  │
│  └─────────────────────┬─────────────────────────────────┘  │
│                        │                                     │
│  ┌─────────────────────┴─────────────────────────────────┐  │
│  │            Tool Router & Safety Layer                  │  │
│  │  • Intent classification                              │  │
│  │  • Safety assessment                                  │  │
│  │  • Approval requests                                  │  │
│  └─────────────────────┬─────────────────────────────────┘  │
│                        │                                     │
│         ┌──────────────┼──────────────┐                     │
│         │              │              │                     │
│  ┌──────▼────┐  ┌──────▼────┐  ┌──────▼────────┐           │
│  │  File Ops │  │   Shell   │  │  Calculator   │  ...      │
│  │   Tools   │  │   Tools   │  │     Tools     │           │
│  └───────────┘  └───────────┘  └───────────────┘           │
│                                                               │
│  ┌───────────────────────────────────────────────────────┐  │
│  │            Error Handler & Retry Logic                 │  │
│  │  • Exponential backoff                                │  │
│  │  • Circuit breaker                                    │  │
│  │  • Graceful degradation                               │  │
│  └───────────────────────────────────────────────────────┘  │
│                                                               │
└─────────────────────────────────────────────────────────────┘

📂 Files¶

• agent_core.py: Core agent implementation
• tool_system.py: Tool registry and execution
• safety_layer.py: Safety checks and approvals
• conversation_manager.py: State and history management
• error_handling.py: Retry logic and recovery
• example_usage.py: Demo scenarios

🚀 Quick Start¶

# Install dependencies
pip install openai aiohttp

# Set API key
export OPENAI_API_KEY="your-key-here"

# Run the complete example
python example_usage.py

💡 Key Features¶

1. Multi-Turn Conversations¶

Like Codex, maintains context across turns:

agent = CodexInspiredAgent()
agent.run("Analyze the codebase")
# Turn 1: Reads files
# Turn 2: Analyzes code
# Turn 3: Generates report

2. Safe Tool Execution¶

Multiple layers of safety:

# Whitelist checking
# Sandboxing simulation
# User approval for dangerous ops
# Output truncation
# Timeout enforcement

3. Intelligent Routing¶

Automatically routes to appropriate tools:

# "What time is it?" → time tool
# "Calculate 2+2" → calculator tool
# "Read main.py" → file tool

4. Error Recovery¶

Handles failures gracefully:

# Automatic retry with backoff
# Fallback to alternative approaches
# User notification of issues

5. Conversation Persistence¶

Saves and resumes sessions:

agent.save_session("session_123.json")
# Later...
agent.load_session("session_123.json")

📊 Example Scenarios¶

Scenario 1: Code Analysis Task¶

query = "Analyze main.py, find bugs, and suggest fixes"

# Turn 1: Read file
# Turn 2: Analyze for bugs (uses analysis tool)
# Turn 3: Generate fixes (uses code generation)
# Turn 4: Present recommendations

Scenario 2: Data Processing¶

query = "Process sales.csv and create a summary report"

# Turn 1: Read CSV (file tool)
# Turn 2: Calculate statistics (calculator tool)
# Turn 3: Generate visualizations (visualization tool)
# Turn 4: Write report (file tool)

Scenario 3: System Administration¶

query = "Check disk usage and clean up old logs"

# Turn 1: Check disk (shell tool)
# Approval requested: "Execute 'du -sh'?"
# Turn 2: Identify old files (shell tool)
# Approval requested: "Delete 500MB of logs?"
# Turn 3: Clean up (shell tool)
# Turn 4: Verify space freed

🔐 Safety Features¶

Command Whitelisting¶

SAFE_COMMANDS = ["ls", "cat", "grep", "find"]
DANGEROUS_COMMANDS = ["rm", "dd", "mkfs", "shutdown"]

Approval Workflow¶

if is_dangerous_operation(tool_call):
    approval = request_user_approval(tool_call)
    if not approval:
        return "Operation denied by user"

Sandboxing¶

# Network disabled
# File access limited to workspace
# Process timeout: 30s
# Output truncated: 10KB max

📈 Performance Considerations¶

Parallel Tool Execution¶

# Some tools run in parallel
results = await asyncio.gather(
    execute_tool("read_file", {"path": "a.py"}),
    execute_tool("read_file", {"path": "b.py"}),
    execute_tool("read_file", {"path": "c.py"}),
)

Context Window Management¶

# Automatically compress old history
if len(history) > MAX_TURNS:
    history = compress_history(history)

Caching¶

# Cache file reads
# Cache tool results
# Cache LLM responses (deterministic queries)

🧪 Testing¶

# Run unit tests
python -m pytest test_agent.py

# Run integration tests
python -m pytest test_integration.py

# Run safety tests
python -m pytest test_safety.py

📚 Learning Objectives¶

After studying this example, you should understand:

1. ✅ How to architect a multi-pattern agent system
2. ✅ State management across async operations
3. ✅ Safety-first design principles
4. ✅ Error handling and recovery strategies
5. ✅ Tool orchestration and routing
6. ✅ User interaction patterns
7. ✅ Performance optimization techniques

🔗 Codex Comparison¶

🚀 Extension Ideas¶

1. Add MCP Support: Integrate with external MCP servers
2. Implement Streaming: Stream tool output in real-time
3. Add Telemetry: OpenTelemetry integration
4. Build TUI: Terminal UI like Codex
5. Enhance Sandboxing: Use Docker or similar
6. Add Planning: Explicit planning tool like Codex
7. Multi-Agent: Coordination between multiple agents

📄 License¶

Educational example for learning purposes.

Start with: example_usage.py to see it in action!