6. Tool Use — Acting on the World

  ┌──────────────────────────────────────────────────────┐
  │  System prompt:                                      │
  │  "You are a helpful assistant."                      │
  │                                                      │
  │  Available tools:                                    │
  │    get_weather(city: string) → weather data          │
  │    web_search(query: string) → search results        │
  ├──────────────────────────────────────────────────────┤
  │  User: "What's the weather in Tokyo?"                │
  └─────────────────────────┬────────────────────────────┘
                            │
                            ▼
                     ┌─────────────┐
                     │     LLM     │
                     └─────────────┘
                            │
            Tool call: get_weather(city="Tokyo")
                            │
                            ▼
                   ┌─────────────────┐
                   │  Application    │
                   │  executes tool  │
                   └────────┬────────┘
                            │
                            ▼
         Result: { temp: "28°C", condition: "humid" }
                            │
                            ▼
                     ┌─────────────┐
                     │     LLM     │
                     └─────────────┘
                            │
                            ▼
        "It's currently 28°C and humid in Tokyo."

Behind the scenes — what the LLM sees at each step:

Step 1:  [system: "You are a helpful assistant.",
          tools: [get_weather, web_search],
          user: "What's the weather in Tokyo?"]
              → LLM responds with: tool_call: get_weather(city="Tokyo")

Step 2:  [system: "...", tools: [...],
          user: "What's the weather in Tokyo?",
          assistant: tool_call: get_weather(city="Tokyo"),
          tool: { temp: "28°C", condition: "humid" }]
              → LLM responds with: "It's currently 28°C and humid in Tokyo."

The LLM doesn't execute tools itself. It only decides which tool to call with which parameters. The application around the LLM performs the actual call and feeds the result back.

Tool definitions are just text in the context window — typically part of the system prompt. The model has learned to recognize this format and generate matching structured calls.

Typical tools: Web search, database queries, API calls, file operations, email access, CRM access, code execution.

Code execution deserves a closer look. Most tools let the model fetch or manipulate information in the outside world — search the web, query a database, send an email. Code execution is different — it lets the model compute, compensating for weaknesses built into how LLMs work.

Why code execution stands out: Remember the strawberry problem from section 1? The model can't count letters in "strawberry" because it never sees individual letters — only tokens. But give the model a code execution tool, and it writes 'strawberry'.count('r') — correct, every time. You can often trigger this simply by asking the model to "use code." The model didn't get smarter. It got a calculator. The same applies to arithmetic, date calculations, sorting, and anything else where precise computation beats pattern matching. This is why the same model gives better answers in a product that has code execution than in one that doesn't.

There's a second reason code execution matters. Ask an LLM to perform a task — say, extract data from a document — and run it 100 times. You'll get dozens of slightly different results, some with small variations, some with large ones. Now ask the LLM to write a small program that performs the same task, and run that 100 times. The results are identical every time. Code execution lets the model move from probabilistic to deterministic — from "roughly right, differently each time" to "exactly right, every time."

Tool integrations are increasingly standardized through protocols like MCP (Model Context Protocol), which aim to make tools portable across different AI systems — define the tool once, use it with any model.