Building a Calculator Tool for AI Agents: Step-by-Step Tutorial

Why Build a Calculator Tool?

LLMs are notoriously unreliable at arithmetic. They can set up equations correctly but frequently miscalculate the result. A calculator tool solves this by offloading the computation to deterministic code. It is also the simplest possible tool to build, making it an ideal starting point for understanding the full tool-calling lifecycle.

This tutorial walks through building a calculator tool, registering it with an agent, and handling the execution loop.

Step 1: Define the Tool Schema

The schema tells the LLM what the tool does and what parameters it accepts:

calculator_schema = {
    "type": "function",
    "function": {
        "name": "calculate",
        "description": "Evaluate a mathematical expression and return the numeric result. Use this for any arithmetic, percentages, or mathematical calculations. Input must be a valid Python math expression.",
        "parameters": {
            "type": "object",
            "properties": {
                "expression": {
                    "type": "string",
                    "description": "A mathematical expression to evaluate, e.g. '(25 * 4) + 17' or '150 * 0.15'. Use Python syntax for operations."
                }
            },
            "required": ["expression"]
        }
    }
}

The description explicitly says "Python math expression" to guide the LLM toward valid syntax like ** for exponents instead of ^.

Step 2: Implement the Tool Function

Never use eval() on untrusted input. Instead, use Python's ast module to parse the expression safely:

import ast
import operator
import math

SAFE_OPERATORS = {
    ast.Add: operator.add,
    ast.Sub: operator.sub,
    ast.Mult: operator.mul,
    ast.Div: operator.truediv,
    ast.FloorDiv: operator.floordiv,
    ast.Mod: operator.mod,
    ast.Pow: operator.pow,
    ast.USub: operator.neg,
    ast.UAdd: operator.pos,
}

SAFE_FUNCTIONS = {
    "sqrt": math.sqrt,
    "abs": abs,
    "round": round,
    "min": min,
    "max": max,
}

def safe_eval(node):
    if isinstance(node, ast.Expression):
        return safe_eval(node.body)
    elif isinstance(node, ast.Constant):
        if isinstance(node.value, (int, float)):
            return node.value
        raise ValueError(f"Unsupported constant: {node.value}")
    elif isinstance(node, ast.BinOp):
        left = safe_eval(node.left)
        right = safe_eval(node.right)
        op_func = SAFE_OPERATORS.get(type(node.op))
        if op_func is None:
            raise ValueError(f"Unsupported operator: {type(node.op).__name__}")
        return op_func(left, right)
    elif isinstance(node, ast.UnaryOp):
        operand = safe_eval(node.operand)
        op_func = SAFE_OPERATORS.get(type(node.op))
        if op_func is None:
            raise ValueError(f"Unsupported unary operator: {type(node.op).__name__}")
        return op_func(operand)
    elif isinstance(node, ast.Call):
        if isinstance(node.func, ast.Name) and node.func.id in SAFE_FUNCTIONS:
            args = [safe_eval(arg) for arg in node.args]
            return SAFE_FUNCTIONS[node.func.id](*args)
        raise ValueError(f"Unsupported function call")
    else:
        raise ValueError(f"Unsupported expression type: {type(node).__name__}")

def calculate(expression: str) -> str:
    try:
        tree = ast.parse(expression, mode="eval")
        result = safe_eval(tree)
        return str(result)
    except (ValueError, SyntaxError, TypeError, ZeroDivisionError) as e:
        return f"Error: {str(e)}"

This evaluator supports basic arithmetic, exponentiation, and a whitelist of safe functions without exposing the system to code injection.

Step 3: Wire It Into the Agent Loop

Here is a complete agent loop using the OpenAI API that calls the calculator tool:

from openai import OpenAI

client = OpenAI()

def run_agent(user_message: str) -> str:
    messages = [
        {"role": "system", "content": "You are a helpful assistant. Use the calculate tool for any math."},
        {"role": "user", "content": user_message}
    ]

    while True:
        response = client.chat.completions.create(
            model="gpt-4o",
            messages=messages,
            tools=[calculator_schema],
        )
        msg = response.choices[0].message
        messages.append(msg)

        if msg.tool_calls:
            for tool_call in msg.tool_calls:
                import json
                args = json.loads(tool_call.function.arguments)
                result = calculate(args["expression"])

                messages.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "content": result,
                })
        else:
            return msg.content

answer = run_agent("What is 15% tip on a $247.50 dinner bill split 3 ways?")
print(answer)

The agent loop continues until the LLM stops making tool calls and returns a text response. Each tool call result is appended with the matching tool_call_id so the LLM can correlate results to requests.

Step 4: Handle Edge Cases

Your calculator will receive unexpected inputs. Build robustness into the tool function:

def calculate(expression: str) -> str:
    if not expression or not expression.strip():
        return "Error: Empty expression"
    if len(expression) > 500:
        return "Error: Expression too long"
    try:
        tree = ast.parse(expression, mode="eval")
        result = safe_eval(tree)
        if isinstance(result, float) and (math.isinf(result) or math.isnan(result)):
            return "Error: Result is infinity or undefined"
        return str(round(result, 10))
    except Exception as e:
        return f"Error: {str(e)}"

Returning a clear error string instead of raising an exception lets the LLM recover by adjusting the expression and trying again.

FAQ

Why not just use Python eval() for the calculator?

Using eval() on LLM-generated strings is a critical security vulnerability. The LLM could produce expressions like __import__('os').system('rm -rf /') either through prompt injection or a malformed response. The AST-based evaluator restricts execution to pure mathematical operations.

Can the LLM call the calculator multiple times in one turn?

Yes. If the model generates multiple tool_calls in a single response, you should execute all of them and return all results. The model might break a complex calculation into steps, calling the calculator for each one.

How do I test that my tool schema works correctly?

Send test prompts that should trigger tool calls and verify the LLM generates valid arguments. Common failure modes include the LLM using ^ for exponents instead of **, or passing expressions with variables. Add these as examples in your tool description to guide correct usage.

---

#ToolBuilding #FunctionCalling #Python #AIAgents #AgenticAI #LearnAI #AIEngineering