When Not to Use an Agent

An agent lets the model choose every step. That costs one model call per step: seven calls for one return. And the model decided nothing surprising on the way, because the steps of a return are the same every time.

When you know the steps before the run starts, write them yourself. That is a workflow: your code decides the steps, and the model does only what code can't do. The tool functions are plain Python functions, so your code can call them without asking a model:

stock = check_stock({"product_id": 3})
if stock["stock"] == 0:
    print("Sold out")

Here an if decides, not a model: it costs nothing and it decides the same way on every run. In this lesson's workflow the model has one job left, writing a friendly reply, in a call without tools and without a loop.

The finished program runs the same return both ways. After the agent's trace it prints:

Workflow:
  Customer is asked: Start the return of order 8?
  Customer answers: yes
  Result: {'id': 8, 'customer': 'Ana Pop', 'ordered': '2026-05-28', 'status': 'returning'}
  Reply: Hi Ana, good news: the return of order 8 has started. We'll write to you when your parcel reaches us.
Agent: calls 7 cost 0.0092
Workflow: calls 1 cost 0.0004

Your task: finish run_workflow: find the order with the product in two loops, call start_return, and ask the model once for the reply. Then make both functions return their calls and cost, and print the comparison, with each cost rounded to 4 decimals.

Assignment
Follow the Coding Tutorial to write the return as a fixed workflow and to compare it with the agent.

Hint
Each tool function takes one dictionary and returns one, so the list you need is under a key of the result: ["orders"] or ["items"]. The workflow's model call has no tools argument.