Function calling

We like to think of function calling as giving AI a phone. She can dial out to real tools, not just talk in the abstract. Instead of just answering, she can actually do something—fetch weather, send a message, check inventory. Suddenly she’s less trivia partner, more coworker.

What it is

Function calling is a bridge. AI maps plain text (“get today’s forecast”) into a specific function with parameters (“weather(city=‘Boston’, day=‘today’)”). From there the function does its job, and she translates the result back into words we can use. Simple, but powerful.

Why it matters

Without this, she’s stuck explaining. With it, she helps finish the task. We don’t want to know the API docs; we want the number that tells us if the train is late. Function calling gives her the keys.

Designing the workflow

Start small. Define a clear function, test it, and see if she calls it correctly. Keep the responses predictable, because unpredictable results make debugging painful. Add functions one at a time. If she can book a table, wait until that works before teaching her to order wine.

Integration choices

Pick the right layer. Sometimes the AI lives inside an app and calls the app’s own functions. Sometimes she sits in the middle, brokering requests between services. Either way, the design rule is the same: functions should be clean and boring. The magic is in her, not the plumbing.

Our takeaway

We spend a lot of time connecting code to other code. Function calling just makes it conversational. We ask, she acts, and the workflow feels less like scripting and more like collaboration. Makes us wonder if, in a year or two, we’ll be designing APIs mainly for her instead of for us.