In plain English
It is the same mechanism now generally called tool calling. The older name describes the shape more precisely: the model outputs a function name and a set of arguments.
The important detail is that the model does not execute anything. It requests. Your code decides whether to run it, which is where every safety control belongs.
What to know
Structured output
A function name and typed arguments, not free text.
The model requests
Execution is your code's decision, not the model's.
Schema-defined
You declare what functions exist and what arguments they take.
Validate everything
Arguments can be wrong or malicious. Treat them as untrusted input.
Why it matters
Knowing the model only requests rather than executes clarifies where responsibility sits. Every guardrail, permission check and validation belongs in your code, not in the prompt.
Common mistakes
×Executing requested calls without validating arguments.
×Vague function descriptions, so the wrong one gets chosen.
×No error handling when a call fails.
×Treating the prompt as the security boundary.
FAQs
Is it different from tool calling?
No. Tool calling is the current term for the same thing.
Who runs the function?
Your code does. The model only asks.
WRITTEN BY TARIQ SALLAM
Marketing Consultant. Entrepreneur. Content Creator.
I'm a marketing consultant, entrepreneur and content creator. I help businesses grow through practical marketing, websites, SEO, content and AI.
More About Tariq →