Function Calling Tool

Define custom functions for agents to extend capabilities with business logic, API calls, and database queries.

Function Calling

Function calling enables agents to invoke custom functions you define, extending their capabilities with business logic, external APIs, and system integrations.

How It Works

1. Define functions with clear descriptions and parameters
2. Agent determines need based on user request
3. Run status changes to requires_action
4. Your code executes the actual function
5. Submit results back to agent
6. Agent processes and generates response

The agent requests function calls but doesn’t execute them. Your application code must handle execution.

Quick Start

Python

from azure.ai.projects.models import FunctionTool
import json

def get_weather(location: str) -> str:
    """
    Get current weather for a location.
    
    :param location: City and state (e.g., "Seattle, WA")
    :return: Weather as JSON string
    """
    # Mock data
    weather_data = {
        "Seattle, WA": "Sunny, 72°F",
        "New York, NY": "Cloudy, 65°F"
    }
    return json.dumps({"weather": weather_data.get(location, "Unknown")})

# Register function
function_tool = FunctionTool(functions={get_weather})

agent = project.agents.create_agent(
    model="gpt-4o",
    name="weather-assistant",
    instructions="Help users check weather using the get_weather function",
    tools=function_tool.definitions,
)

Handling Function Calls

Python

run = project.agents.runs.create(thread_id=thread.id, agent_id=agent.id)

while run.status in ["queued", "in_progress", "requires_action"]:
    if run.status == "requires_action":
        tool_outputs = []
        for tool_call in run.required_action.submit_tool_outputs.tool_calls:
            if tool_call.function.name == "get_weather":
                args = json.loads(tool_call.function.arguments)
                output = get_weather(args["location"])
                tool_outputs.append({
                    "tool_call_id": tool_call.id,
                    "output": output
                })
        
        project.agents.runs.submit_tool_outputs(
            thread_id=thread.id,
            run_id=run.id,
            tool_outputs=tool_outputs
        )
    
    run = project.agents.runs.get(thread_id=thread.id, run_id=run.id)

Best Practices

• Write clear function descriptions
• Define required parameters explicitly
• Return structured JSON
• Handle errors gracefully
• Implement timeout logic (runs expire after 10 minutes)

See Azure Functions for serverless hosting.