How AI Takes Shortcuts: Core Skills of Large-Model Autonomous Decision-Making via Tool Calls

Large language models (LLMs) are fundamentally language predictors; they excel at generating text but lack up‑to‑date knowledge and arithmetic accuracy. To overcome these limits, LLMs can invoke external tools (Tool Use) such as APIs, code modules, or databases, turning them into executable agents.

Why Tool Use Is Needed

Capability boundary: LLMs cannot perform precise calculations (e.g., GPT‑3’s 99×99 error rate reaches 76%).

Data freshness: After training, a model’s knowledge is frozen and cannot reflect real‑time information like today’s weather.

By providing a list of tools to the model, it can decide when to call them and incorporate the results into its responses.

Three‑Stage Workflow

The process consists of:

Function Call (instruction generation): The model outputs a structured JSON containing the tool name and parameters.

Tool Use (execution): An external system parses the JSON, invokes the corresponding API or code, and returns the result.

Result Integration: The model formats the tool’s output into natural language for the user.

Example request format:

{
  "messages": [{"role": "user", "content": "圆周率平方是多少？"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "math_api",
      "description": "数学计算服务",
      "parameters": {
        "type": "object",
        "properties": {"expression": {"type": "string"}}
      }
    }
  }]
}

When the model decides a tool is needed, it returns:

{
  "tool_calls": [{
    "id": "call_abc123",
    "type": "function",
    "function": {"name": "math_api", "arguments": "{\"expression\": \"π^2\"}"}
  }]
}

Autonomous Routing and Parameter Generation

Instead of hard‑coded if‑else branches, the model can automatically choose the appropriate tool based on the user query:

# Pseudocode example
if "天气" in query:
    call weather_api()
elif "计算" in query:
    call calculator()
else:
    generate_normal_response()

For complex requests, the model extracts parameters automatically, e.g., from “帮我在静安寺附近找人均200的意大利餐厅” it produces:

{
  "location": "静安寺",
  "price_range": 200,
  "cuisine": "意大利菜"
}

Error Handling

If a tool call fails, the model receives an error payload and can respond gracefully:

{
  "role": "tool",
  "content": "{\"error\": \"服务超时\"}",
  "tool_call_id": "call_123"
}

Function Call vs. Tool Use

Function Call is the standardized instruction that tells an external system what to execute; Tool Use is the actual execution of that instruction. Function Call serves as the precondition for Tool Use, while Tool Use realizes the task.

Layered Implementation

The architecture separates the two stages for flexibility and security. The model never directly accesses resources; the sandboxed tool platform isolates execution, preventing unsafe operations.

Illustrative Flow

User asks, “北京今天气温多少？” The model generates a JSON tool call (Function Call), the system invokes the weather API (Tool Use), and the model returns a natural‑language answer such as “北京今天25℃，晴。”

In summary, tool calls empower LLMs to overcome their inherent limitations by delegating tasks to external services, enabling real‑time data retrieval, precise computation, and multi‑tool coordination while keeping the model’s core safe and isolated.