function-tool-author

function-tool-author

Author a Python function tool for ADK 2.0. ADK reads type hints and the docstring to build the tool schema sent to the LLM.

When to use

• Wrap a deterministic capability (math, lookups, transforms)
• Bridge to your internal services (DB, API, filesystem)
• Anything where you don't want LLM-generated code/output

Template

from typing import Optional

def get_weather(city: str, units: str = "fahrenheit") -> dict:
    """Get the current weather for a city.

    Args:
        city: City name (e.g., 'San Francisco', 'Tokyo').
        units: Temperature units. One of 'fahrenheit' or 'celsius'.

    Returns:
        A dict with keys: temperature (number), conditions (string),
        humidity (number, percent).
    """
    # Replace with your real call
    return {
        "temperature": 68,
        "conditions": "partly cloudy",
        "humidity": 55,
        "city": city,
        "units": units,
    }

Bind to an agent

from google.adk.agents import LlmAgent

root_agent = LlmAgent(
    name="weather_assistant",
    model="gemini-2.5-flash",
    instruction="Answer weather questions using get_weather.",
    tools=[get_weather],
)

Schema rules ADK enforces

• Type hints required on all parameters and return type
• Docstring required with Args: block — these become parameter descriptions
• Supported types: str, int, float, bool, list[T], dict, Optional[T]
• Avoid complex Pydantic models in arguments — flatten to primitives where possible
• Return JSON-serializable values

Async tools

async def fetch_user(user_id: str) -> dict:
    """Fetch a user record by ID.

    Args:
        user_id: The unique user identifier.

    Returns:
        User record with name, email, created_at.
    """
    async with httpx.AsyncClient() as client:
        r = await client.get(f"https://api.example.com/users/{user_id}")
        return r.json()

ADK auto-detects async functions; bind same way as sync.

Error handling

def divide(a: float, b: float) -> dict:
    """Divide a by b. Returns {'result': float} or {'error': str}."""
    if b == 0:
        return {"error": "Cannot divide by zero"}
    return {"result": a / b}

Return errors as data — don't raise. The LLM can recover from a structured error.

Validation

• Function imports cleanly (no module-level side effects)
• Type hints + docstring complete
• Manual smoke test: print(get_weather("Tokyo")) works
• Bind to agent and ask via adk web — ADK shows the function call in the UI