Is a chatbot with tools an agent?

It depends on who decides the next step. If the model answers your message, perhaps calling a tool once, and then waits for you, you still drive every turn and the system behaves like a chatbot. It becomes an agent when the model keeps choosing tools, reading the results and deciding what to do next on its own until a goal or a stopping rule ends the run: it runs tools in a loop to achieve a goal. Chapter 1 (free) draws the line with three worked systems.

If my pipeline has a router, is it an agent?

No. A router is a workflow that grants the model one decision: it picks which hardcoded branch handles the request. Every branch was drawn in advance, so the flowchart exists before the request arrives, which is the book's litmus test for a workflow. On the autonomy dial it sits one stop above the fixed pipeline and well short of an agent with approval gates; Chapter 1 walks the whole dial.

Why does the book care so much about who owns the control flow?

Because moving the control flow from your code to the model is the one change that makes agents useful, and also what makes them, in Chapter 1's words, "expensive, slow, and occasionally alarming." A chatbot's worst output is a wrong answer; an agent's mistake is a wrong action, followed by more actions built on top of it. Knowing who decides the next step tells you where verification, budgets and approval gates have to go.

Do I even need an agent?

Often not. The book's default is a ladder: plain code, then a single model call, then a workflow, then an agent, and the discipline is to stay on the lowest rung that solves the problem. Each rung up buys adaptability and pays for it in money, latency and predictability. The last section of Chapter 1 prices the climb, and the Should this be an agent? tool walks you through the decision.

Learn / Explainers / What is an agent? Chatbot, workflow, and agent

What is an agent? Chatbot, workflow, and agent

Animated explainer · 3:28 · from Chapter 1

Space plays and pauses · ← → step · F fullscreen

A three-minute animated explainer: chatbot, workflow and agent differ in one thing, who decides the next step. Then the litmus test and the autonomy dial.

Transcript

Every step of the animation, in the words it uses on screen. Select a line to jump to it.

From Chapter 1

What Is an Agent?

This explainer condenses a few pages of Chapter 1. The chapter itself traces the ideas through a worked example and is free to read online.

Read Chapter 1 free See the formats

Questions

Is a chatbot with tools an agent?: It depends on who decides the next step. If the model answers your message, perhaps calling a tool once, and then waits for you, you still drive every turn and the system behaves like a chatbot. It becomes an agent when the model keeps choosing tools, reading the results and deciding what to do next on its own until a goal or a stopping rule ends the run: it runs tools in a loop to achieve a goal. Chapter 1 (free) draws the line with three worked systems.
If my pipeline has a router, is it an agent?: No. A router is a workflow that grants the model one decision: it picks which hardcoded branch handles the request. Every branch was drawn in advance, so the flowchart exists before the request arrives, which is the book's litmus test for a workflow. On the autonomy dial it sits one stop above the fixed pipeline and well short of an agent with approval gates; Chapter 1 walks the whole dial.
Why does the book care so much about who owns the control flow?: Because moving the control flow from your code to the model is the one change that makes agents useful, and also what makes them, in Chapter 1's words, "expensive, slow, and occasionally alarming." A chatbot's worst output is a wrong answer; an agent's mistake is a wrong action, followed by more actions built on top of it. Knowing who decides the next step tells you where verification, budgets and approval gates have to go.
Do I even need an agent?: Often not. The book's default is a ladder: plain code, then a single model call, then a workflow, then an agent, and the discipline is to stay on the lowest rung that solves the problem. Each rung up buys adaptability and pays for it in money, latency and predictability. The last section of Chapter 1 prices the climb, and the Should this be an agent? tool walks you through the decision.