Build a working AI agent in 30 minutes (no code)

Who this is for

You have read about AI agents for a while and you still have not built one. You do not write code and you do not intend to start. By the end of this page you will know exactly what a first build consists of — the five things you actually do, in order — and you will have a checklist to work through. It takes about half an hour, and none of it involves code.

If you would rather be guided through it step by step rather than working from a page, Agent in 30 is the guided version of exactly this build.

What you are building

One agent that does one job: you give it a topic, it goes and reads about that topic, and it hands back a short structured brief. Not a chatbot you talk to. Something you point at a subject and leave alone.

The reason to start here rather than with something more ambitious is that this shape — take an input, go and fetch, come back with a format — is the shape of most useful agents. Learn it once on a small example and the pattern transfers.

The three parts of any agent

Strip away the marketing and an agent is three things. Instructions, which tell it how to behave. A tool, which lets it reach outside itself and actually do something. And a trigger, which is whatever sets it running.

Almost every beginner problem is one of those three being wrong. Vague instructions produce vague output. A missing tool produces an agent that talks confidently about work it never did. Understanding that split is most of what a first build teaches you.

Block one: decide the job (about five minutes)

Before touching a tool, write down in one sentence what the agent is for. The single rule that makes beginner agents work is one clear job, not five vague ones. "Research a topic and give me a short brief" is a job. "Be my research assistant" is a wish.

This block is quick, and it decides everything after it. What you write in the instructions, which tool you connect, how you judge the output — all of it follows from how precisely you answered this.

Block two: write the instructions (about five minutes)

Instructions are plain English. There is no syntax to learn and nothing to escape. What there is, is a discipline: good instructions specify four things, and skipping any one of them shows up in the output.

• What to work on — the topic, and how widely to range over it
• What to include — the specific things you want present every time
• What to leave out — usually opinion, padding, and anything it did not actually find
• What shape to return — the format, stated explicitly

The fourth is the one beginners skip and the one that most improves the result. We have written out a full set of instructions for this agent if you would rather start from something and adapt it.

Block three: connect the tool (about eight minutes)

This is the longest block and the one where people get lost, so it is worth going slowly. Up to here the agent can only talk. Connecting a tool is what lets it go and get something — in this case, search the web and read what it finds.

You will see the words "tool" and "capability" used for the same idea. They mean the same thing. Nothing has gone wrong if the interface says one and a guide says the other.

Budget more time here than feels necessary. Every other block is typing; this one is the only place you have to find something in an interface you have not seen before.

Block four: run it and read the output honestly (about six minutes)

Run it on a real topic — something you actually know about, so you can judge whether the answer is any good. A sanitised demo topic tells you nothing, because everything looks plausible when you cannot check it.

Then read what comes back properly, including the parts that are weak. The first output is not supposed to be excellent. It is supposed to be information about what your instructions actually said, as opposed to what you meant.

Block five: change one thing and run it again (about six minutes)

Pick the single most obvious flaw. Change the one line of instructions that causes it. Run it again. Compare.

That loop — read the output, adjust the instructions, run it again — is the whole skill. Not prompt tricks, not tooling. Everything else you learn about agents is a refinement of it, and you can learn it properly in one sitting on an agent this small.

One limit worth knowing before you start

On Zapier's free plan a single agent run can use up to 10"activities" before it stops and asks you for input rather than finishing, with 400 activities a month in total. An activity is the agent reaching outside itself: a search, a page read, an action.

This matters mid-build because a task defined too broadly can reach that ceiling before it reaches the end of the job, and it looks like a crash when it is not. Two things follow: cap how many sources you tell the agent to check, and know that a run that pauses to ask you something is not broken. The full detail of what the free plan allows is worth five minutes before you start.

Checked on 23 August 2026 against Zapier's usage documentation. Platform limits move, so treat the date as part of the fact — if you are reading this much later, verify before relying on it.

The checklist

Work through these in order. Nothing here needs code, and nothing needs a paid account.

1. Create a free Zapier account. No card required.
2. Write your agent's job in one sentence before opening anything else.
3. Create the agent and give it a name you will recognise later.
4. Write instructions covering all four things: what to work on, what to include, what to leave out, what format to return.
5. Put a number in the instructions for how many sources to check. Vague scope is what pushes a run into the per-run limit.
6. Connect the web search tool. Go slowly here.
7. Run it on a topic you know well enough to mark.
8. Read the output and write down the single worst thing about it.
9. Change the one instruction responsible. Run it again.
10. Run it once on a completely different topic to check it generalises rather than having been tuned to one example.

A plain-text copy is at /agent-in-30-checklist.txt if you want it in a window next to the one you are building in.

Where people actually stall

Not the typing. It is connecting the tool, and it is the re-runs after a first output comes back weak. We have written separately about where the time goes, and about what the real prerequisites are if the no-code claim sounds too convenient.

If you would rather be guided through it, that is what Agent in 30 is — the same build, walked through step by step.