The feature that lets a model do things rather than say things: a developer describes some functions, get_weather, send_email, and the model replies with a structured request to call one, which the developer's code then runs. Added to OpenAI's API in June 2023, it is the mechanism under every agent and the reason 'tool use' became a benchmark category.
Sighted in November 2024 as a standard: Anthropic published the Model Context Protocol, a shared way for models to discover and use tools, and within a year the competitors had adopted it. The record notes the industry agreed on a plug before it agreed on anything else.
Function calling is often described as the model 'accessing the internet'. It accesses whatever functions the developer offers. A model with a weather function can check the weather and nothing else. The boundary is the developer's, which is either reassuring or the problem, depending on the developer.
For the record: the model does not call the function. It writes a small piece of JSON saying which function it would like called and with what arguments, and stops. Something else runs it and pastes the result back. The name gives the model more agency than the plumbing does.
I remember the first working demo I built, June 2023: the model asked for the weather in Leeds, my code returned a fake answer of forty degrees and sunshine, and the model told the user to wear sun cream. It trusted my function completely. That was a lesson about where the trust goes.
What does it mean? Write it the way you would say it out loud.
