The Line Cook's Tools

Opening Scene

A line cook without actual knives, a working stove, and real ingredients can describe a dish in perfect, vivid detail and still never actually produce it. Description isn’t the same as capability. An AI agent faces this exact same gap: it can generate perfectly coherent text describing what it would do, and that text accomplishes literally nothing in the real world unless the agent has actual tools it can invoke to take genuine action.

In Plain English

Tool use (or function calling) lets an agent invoke actual external capabilities — searching the web, querying a database, sending an email, running code, calling an API — rather than simply generating text about what should happen. This relies directly on the structured output techniques covered in this content library’s dedicated prompt engineering series: the model generates a structured request specifying which tool to call and with what parameters, an external system actually executes that call, and the result gets returned to the agent to inform its next step.

The Old Way

Before reliable tool use was practical, language models were fundamentally limited to generating text with no way to affect anything outside that text:

• Early language models could only generate text, with no built-in mechanism for invoking an external system or taking a genuine action.
• Connecting a model’s output to a real action required significant custom engineering, parsing free-form text output to try to infer an intended action — a fragile, error-prone approach.
• Without reliable structured output, covered in this content library’s prompt engineering series, translating a model’s intent into a genuine, correctly formatted function call was unreliable.

Dedicated function calling capabilities emerged specifically to make this translation from intent to action reliable, rather than a fragile, best-effort text-parsing exercise.

What’s Changing (and Why AI Is the Reason)

1. Most modern LLM APIs now offer dedicated function calling features, letting a model reliably specify which tool to invoke and with what exact parameters, connecting directly to the structured output techniques covered in this content library’s prompt engineering series.
2. A growing ecosystem of standardized tool integrations has emerged, making it considerably easier to connect an agent to real-world systems — databases, APIs, file systems — without custom, one-off engineering for each connection.
3. This capability is precisely what separates a genuine agent from a simple chatbot: the ability to actually change something in the world, not just describe what should change.

The Metaphor, Fully Extended

The Kitchen	Tool Use Concept
A cook describing a dish perfectly without any actual knives or stove	A model generating text describing an action with no way to execute it
Real knives, a working stove, and actual ingredients	Real, invokable tools like search, database queries, and APIs
A structured order sent to a specific station with exact specifications	A structured function call sent to a specific tool with exact parameters
A dish that actually gets produced and served	An action that actually gets taken and its result returned

Quick Recap

• Tool use lets an agent invoke actual external capabilities, rather than just generating text describing an action.
• This relies on reliable structured output, connecting directly to this content library’s prompt engineering series.
• Dedicated function calling features and a growing tool integration ecosystem have made this considerably more accessible.
• Tool use is the specific capability that separates a genuine agent from a simple chatbot.

Where This Fits in the Series

Article 6 covered giving an agent real tools to work with. Article 7 covers what happens when a single agent’s own tools aren’t enough for a task, and it needs to call for backup.