The tool loop

messages = [user question]
for step in range(max_steps):
    reply = model(messages, tool_schemas)
    if no tool calls:  return the text          # done
    append the assistant message (with its tool calls)
    for each call:     run it, append a tool message with the call's id and the result
return "stopped: too many steps"

Details that must be exactly right, or providers will usually reject the request:

1. Append the assistant message that contains the tool calls, before the results.
2. Every call gets a result message with the same id. The order may differ, but none may be missing. A test checks that the ids match one to one.
3. A model may request several calls at once. The example asks for the weather and a sum together. They are independent, so you may run them concurrently (Level 0, module 4).
4. Give the model the results and call it again. Only a reply with no tool calls is a final answer.
5. Cap the number of steps. A test uses a model that never stops asking. The loop ends after max_steps and reports stopped = "max_steps" instead of spending money forever.

A step cap is the simplest stop. Real agents also cap total tokens, cost and elapsed time (module 6, and Level 11).