6.1 What is Tool Use? — LLMs Can Delegate

🎯 Core Goals

Understand that LLMs can call external software to solve their limitations.
See the power of “delegation” for math, search, and data access.

LLMs are great at language but bad at math and facts. Tool Use lets them delegate to a calculator or a search engine instead of guessing.

👁️ Visuals & Interactives

LLMs Can Delegate

See how an LLM uses a calculator to avoid "guessing" math.

Without Tools

Question: "What is 123 × 456?"

AI Logic: "I remember seeing numbers like this. It's probably around 50,000..."

Output: "123 × 456 is 56,128." (Wrong! It guessed.)

handyman Tool Call

With Calculator Tool

Question: "What is 123 × 456?"

CALL: calculator.multiply(123, 456)
RESULT: 56088

Output: "123 × 456 is 56,088." (Correct! Exact calculation.)

By delegating to a calculator, the LLM stops being a "statistical guesser" and becomes a precise processor.

But wait — how does the LLM actually "use" a tool?

Step 1 — LLM Output

I need to calculate this.
[tool_call: calculator.multiply(123, 456)]
Let me use the result...

The LLM only outputs text — including the trigger phrase. It never "runs" anything itself.

arrow_forward

Step 2 — Executor Intercepts

Detected: tool_call
Running: calculator.multiply(123, 456)
Result: 56088 ✓

A separate program reads the trigger, executes the real action, and sends the result back.

🎮 Think of it as Simon Says: the LLM is Simon — it only says the command. Someone else actually carries it out.

🎮 The Simon Says Trick

Here’s the key insight most people miss: an LLM never actually “does” anything.

It only ever outputs text. That’s it. No clicking, no downloading, no running code — just text.

So how does it “use” a calculator? Think of the game Simon Says. When Simon says something, someone else carries out the action. The LLM is Simon — it only says the instruction. A separate program listens for the trigger and actually executes the task.

In practice, it looks like this:

LLM outputs: [tool_call: calculator.multiply(123, 456)]

That’s just text with a special trigger pattern. Another program reads it, runs the calculation, and feeds the result back. The LLM then reads that result and continues writing.

“Simon says, download data from example.com” — the LLM says it; the executor does it.

📝 Key Concepts

Delegation: The LLM decides WHICH tool to use and WHAT to ask — but the tool does the heavy lifting.
Common Tools: Calculators, web search, code execution, file reading, and database queries.
Trigger Text: Tool calls are just specially formatted text that a surrounding system intercepts and runs.
Incorporating Results: Once the tool finishes, the LLM reads the result and incorporates it back into your chat.

What is an API? An API (Application Programming Interface) is like a restaurant menu. You don’t go into the kitchen — you look at the menu, place an order, and get food back. An API is the “menu” that lets software programs talk to each other.

🧠 QUIZ

When an LLM "uses a calculator," what actually happens?

The LLM performs the math internally using a built-in module

The LLM outputs specially formatted text that an external system intercepts and executes

The LLM sends the calculation to another AI model that specializes in math