Agent Capabilities

LLMs can't access the internet, databases, or filesystems natively. Tool calling is the bridge — it gives the LLM a typed API to the real world. Every production agent is only as capable as its tool set.

Tool calling is like giving a remote worker a set of authorized system logins — they can't physically touch the servers, but they can operate them through defined interfaces.

Scenario

Agent needs weather data. LLM returns: tool_call(name='get_weather', args={city: 'Tokyo'}). Runtime calls the weather API, returns {temp: 28, condition: 'sunny'}. LLM uses this to answer the user.

Solution

In Agent Foundations, apply Agent Capabilities to this scenario: Agent needs weather data. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Agent Capabilities (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Agent Capabilities happens in the code.

1from pydantic import BaseModel, Field  # import dependencies2
3class WeatherInput(BaseModel):  # define a data structure or component4    city: str = Field(description="City name")5
6def get_weather(city: str) -> dict:  # define a reusable function7    # calls weather API8    return {"temp": 28, "condition": "sunny", "city": city}  # return the result9
10tools = [{11    "type": "function",12    "function": {13        "name": "get_weather",14        "description": "Get current weather for a city",15        "parameters": WeatherInput.model_json_schema(),16    }17}]18
19response = client.chat.completions.create(  # call the API20    model="gpt-4o",21    messages=[{"role": "user", "content": "Weather in Tokyo?"}],22    tools=tools,23)24# Parse: response.choices[0].message.tool_calls[0].function

• Vague tool descriptions causing wrong tool selection
• No argument validation — injection and type errors
• Unrestricted shell/code execution tools
• No timeout — hung tools block the agent loop
• Prompt-based tool parsing instead of native calling

Quick recap — the most important points from this module.