README.md

Order Processing with Structured Output

Demonstrates order processing with:

• Structured output from LLM agent (action, action_details, action_requested)
• Conditional fraud review for orders > $1000
• Human-in-the-loop suspend/resume

Flow

Amount <= $1000:

1. Agent charges customer via Stripe
2. Agent immediately sends confirmation email
3. Done

Amount > $1000:

1. Agent charges customer via Stripe
2. Agent returns action_requested: "fraud_review"
3. Workflow suspends for human review
4. If approved: Agent sends confirmation email
5. If rejected: Order cancelled

Run

# Terminal 1: Start worker
python worker.py

# Terminal 2: Run workflow
python main.py
# Enter amount: 500 (no fraud review) or 1500 (fraud review)

Approve/Reject Orders

When a workflow suspends for fraud review, use the utility:

# Approve
python approve_order.py <execution_id>

# Reject
python approve_order.py <execution_id> --reject

The execution ID is printed in the worker output when fraud review is required.

Files

• schemas.py - Structured output schema (OrderAgentOutput)
• agents.py - LLM agent with output_schema
• tools.py - charge_stripe, send_confirmation_email
• workflows.py - Order processing with conditional fraud review
• approve_order.py - Utility to approve/reject pending orders