End-to-End Agent QA

Where we left off

No new mechanism this lesson, ask() is Lesson 7's loop, unchanged. What's new is the test: three questions, one needing retrieval, one needing the time tool, one needing neither, run through the exact same function, to see all three decisions the Beginner tier has built land correctly in a single run.

The code, piece by piece

The only change from Lesson 7 is ask()'s return type, now tuple[str, str], so main() can print which tool (if any) actually ran alongside each answer:

def ask(query: str, store: list[dict]) -> tuple[str, str]:
    ...
    if not calls:
        return response.text or "", "none"
    ...
    return final_response.text or "", call.name

Nothing else moved. This lesson exists to demonstrate a property, not introduce one: the same unmodified code correctly handles all three cases because the branching lives inside Gemini's own judgment, not in this course's if statements.

Why this is the Beginner tier's actual milestone

Compare this to what would be required to get the same three answers correctly with the fixed pipeline from Lesson 2: you'd need to write code that first classifies each incoming question as "needs retrieval," "needs the time," or "needs neither," before deciding what to do, essentially reimplementing the judgment Gemini's function calling already gives you as a side effect of asking it a question with tools declared. That classification step is real engineering work in a fixed pipeline; here, it's not code at all.