What an AI agent actually is

That loop — "let the model decide, step by step, whether it needs more information or needs to take an action before it can answer" — is the actual mechanical difference between a chatbot that just talks and an agent that can do things. Everything else people mean by "AI agent" (autonomy, planning, memory, personality) is built on top of this same loop; the loop itself is the part worth understanding cold before any of the fancier framing.

A tool, from the model's side, is just a described function: a name, a natural-language description of what it does, and a schema for its arguments. The model never actually executes anything — it emits a structured request ("call get_available_slots with { date: '2026-06-10' }"), and the application's own code is what actually runs that function and returns a result back into the conversation.

This is the mechanism that lets an agent answer with facts it was never trained on and couldn't possibly know — today's actual appointment availability, a specific customer's specific order status — by handing it the ability to ask for that information mid-conversation, rather than trying to bake it into the model itself.

Generating a response token-by-token takes real, human-perceptible time — often several seconds for a substantive answer. Streaming means sending each token to the client as soon as it's generated, rather than waiting for the entire response to finish and sending it all at once. The mechanism is typically Server-Sent Events (SSE): a long-lived HTTP response the server keeps writing chunks into, that the browser reads incrementally.

Users have been trained by every major chat product to expect the second experience — the visible "typing" effect isn't decorative, it's what makes a multi-second response feel responsive instead of frozen. A non-streaming agent endpoint is a noticeably worse product experience for the exact same underlying model and latency.

Not every agent needs conversation memory, tools, or persistence. A simple product-info chat widget — no login, no history kept between page loads, answering purely from context stuffed into the system prompt (product descriptions, FAQ content) — is a real, complete agent in the narrow sense above, just one where the loop typically never needs to call a tool at all. It's worth building or recognizing this simpler shape first, before reaching for tools, persisted sessions, or write actions, all covered in later lessons in this domain.