Multi-Step Loops

Where we left off

Every Beginner-tier lesson made at most one round trip: model call, tool call if requested, one more model call, done. That's enough for a single-fact question, but not for a compound one needing facts from two different documents. This lesson replaces the if-based single round trip with a while loop that keeps going as long as the model keeps asking for tool calls.

The misconception this lesson corrects

It's easy to read Lesson 5's loop and assume "the agent loop" is inherently a one-shot thing: ask, maybe call a tool, answer. That's an artifact of the questions used so far, not a limit of function calling itself. Nothing about the API caps the number of round trips; Lesson 5's if not calls was simply never exercised more than once. This lesson's question is deliberately compound, needing the sourdough note and the aquarium note, so a single search_notes() call genuinely cannot answer it.

The code, piece by piece

def search_notes(query: str, store: list[dict], k: int = 1) -> str:

k drops from 2 to 1 this lesson, on purpose. With k=2, a single call might accidentally surface both relevant notes anyway (there are only five in the whole store), hiding the multi-step behavior this lesson is trying to demonstrate. k=1 makes each call genuinely single-document, so a two-part question forces two calls.

while True:
    response = client.models.generate_content(model=CHAT_MODEL, contents=contents, config=CONFIG)
    calls = response.function_calls
    if not calls:
        return response.text or "", steps
    ...

The structural change from every earlier lesson: what was an if is now a while True, with the same body, appending the model's call turn and the tool's result turn to contents each iteration, looping back to call generate_content() again with the growing transcript. The loop only exits when a turn comes back with no function call, meaning the model judged it now has enough information to answer.

Why the description had to change too

SEARCH_NOTES_DECLARATION's description now explicitly says results come from "ONE source document at a time" and invites another call for a different query. This isn't decoration, without it the model has no particular reason to call the tool twice instead of assuming one call's result is all there is; Lesson 11 studies this kind of instruction (steering the model toward decomposing a compound question) directly.