Should I build an agent? Three questions, four answers
By Christophe Paka · Facts checked on 23 October 2026
Your manager wants an agent on the roadmap. A client asks for one by name. Before anyone writes code, you need an answer to a plain question: should I build an agent, or something simpler?
Skip the question and you can lose weeks. One freelancer wrote on Reddit about three weeks spent building an agent when a simple workflow would have done the job. An agent is a program where the AI model picks the next step. The figure turns the choice into three questions and four answers. Three of the four are cheaper.
How to read this figure
Start at question 1, top left. Can you write the steps down, start to finish? A note above the box says start with the job, not the tool. A yes sends you down. A no sends you right, along the spine toward the agent.
Take the yes branch down. Is it one step, or several? One step leads to the orange one prompt box: one model call, plus retrieval and good examples. Retrieval means fetching the right documents into the prompt. Several steps lead to the purple workflow box: you write the code path, and the model fills the blanks.
Take the no branch to question 2. Can you check whether a run came out right? Something other than a person reading every run must tell a good result from a bad one. A no drops down to the red box: do not automate this yet.
Go on to question 3. Can you bound the worst thing it could do? To bound means to limit the damage: what the agent can reach, spend or delete. A no also drops to do not automate this yet. A yes leads to the orange agent box at top right: the model picks the next step, inside your limits.
Read the line at the bottom. Three of the four answers are cheaper, faster and easier to debug than the fourth. So dropping off the spine at any question is a win, not a failure. The agent is the last leaf, not the first.
Key takeaways
- Ask the three questions in order, and stop at the first answer that fits.
- Write the step list first; if it holds, ship a workflow.
- Name the check and the bound before you approve an agent; with either missing, say not yet.
Worked example: four jobs through the tree
The book walks real jobs through the three questions. Each one lands on a different leaf.
- Draft a reply to a customer email. You can write the steps: fetch the policy, then draft the reply with it in the prompt. One step: one prompt.
- Add a rule: the reply may quote only refund windows the policy allows. Your code checks the draft and sends it back if it differs. Now there are several steps: a workflow.
- Fix a failing test in a code project. You cannot write the steps, but the project's tests check every try. Bound it with a branch (a separate copy of the code), a container (a sealed box that runs a program apart from the rest of the machine), and a human review. Three answers in favor: agent.
- Pay the invoices in a finance inbox. No cheap check proves an invoice is real, and a wrong payment sends money to a stranger for good. Not yet: keep a human deciding, and automate around them.
Result: Four jobs, four leaves. To turn not yet into yes, build the check and the bound first. See the five workflow patterns for the shapes the workflow leaf can take.
Questions readers ask
What is the point where you decide this should be an agent instead of a normal workflow?
At question 1, the moment the step list breaks. If you can write the steps, code them: a workflow is cheaper and repeatable. If you cannot, an agent is on the table, but only once you can also check a run and bound the damage.
How can I start making AI agents without any background in programming?
For a toy, a chatbot can build one for you. But once the agent sends, buys, deletes or writes, the three questions apply, with code or without. A chatbot builds the agent you ask for. It will not check the result or limit the reach for you.
If you were starting from zero, how would you learn to build an AI agent?
Start where builders tell beginners to start: avoid agents at first. Take one real task, run it through the three questions, and build the leaf it names. Most of the time that is one prompt or a workflow, and you learn the parts an agent is made of.
Sources: Building effective agents (Anthropic); A practical guide to building agents (OpenAI, PDF).
This figure is from AI Agents Made Crystal Clear, chapter 3.
The chapter builds the idea step by step, with recipes you can follow, a Try it to check yourself, and every source.
Get the book
