The Model Doesn't Decide—The Code Does

One insight from the lesson that stuck with me: what we call an AI "agent" is usually not a single autonomous entity. It's a series of prompts orchestrated by traditional code. The model itself—when you call it—is just a function: prompt in, text out. The orchestration layer (the regular software wrapping those calls) is what evaluates the output and decides what prompt to send next based on human-defined conditions.

This matters because it changes how we think about responsibility, testing, and what's actually "happening" in the system. When I understand that the model isn't choosing the next step, I see the control differently. It shifts where I'd look to audit or improve the system.

What does this distinction change for you?

3 comments

The Model Doesn't Decide—The Code Does