Multi-Tool Agents

Where we left off

Every lesson so far declared exactly one tool. This lesson adds a second, deliberately unrelated one, get_current_datetime(), a timezone lookup using Python's standard-library zoneinfo, no new dependency. The point isn't the time tool itself, it's forcing the loop to actually dispatch: given a call the model made, which of several real functions does it correspond to?

The code, piece by piece

def get_current_datetime(timezone: str) -> str:
    try:
        tz = ZoneInfo(timezone)
    except Exception:
        return f"Unknown IANA timezone name: {timezone!r}"
    now = datetime.now(tz)
    return now.strftime("%Y-%m-%d %H:%M %Z")

Chosen specifically because it needs no new dependency (zoneinfo is in the Python standard library) and has nothing to do with retrieval, a clean second tool to prove the loop generalizes.

TOOLS = types.Tool(
    function_declarations=[SEARCH_NOTES_DECLARATION, GET_CURRENT_DATETIME_DECLARATION]
)

One Tool can carry more than one FunctionDeclaration; Gemini sees both and picks (at most) one per turn based on which one's description best matches the question, exactly the same judgment Lesson 6 already showed for the single-tool case, now with a real second option instead of just "answer directly."

def run_tool(call: types.FunctionCall, store: list[dict]) -> str:
    if call.name == "search_notes":
        return search_notes(call.args["query"], store)
    elif call.name == "get_current_datetime":
        return get_current_datetime(call.args["timezone"])
    else:
        raise ValueError(f"Unknown tool: {call.name}")

This is the new piece: a dispatch step between "the model requested a call" and "run the matching Python function." With two tools it's a two-branch if/elif, entirely readable. Hold onto that feeling, this exact shape is what Lesson 19 revisits once a third and fourth tool make it noticeably less pleasant.