A research-and-summary agent you can run today

Who this is for

You want a first agent that does something genuinely useful rather than a demo. This page gives you the job, the instructions to paste, and the reasoning behind each line so you can adapt it. No code, and it runs on a free Zapier account. About half an hour end to end.

The job

You give it a topic. It searches, reads a handful of sources, and returns a short brief in a fixed format. You can run it again on anything.

The reason this is a good first agent is that you can tell immediately whether it worked. Run it on something you already know about and the failures are obvious. Most beginner agents are built on tasks where every output looks equally plausible, which teaches you nothing.

The instructions

Paste this into the agent's instructions field, then read the section below on why each part is there — that is the bit that lets you change it for your own use.

You research topics and return short, structured briefs. You are writing for someone who needs to understand a subject quickly and will act on what you tell them.

When I give you a topic, search for it and read the most relevant sources you find. Check three sources unless I say otherwise. Stop once you have read them, even if you think another would help.

Include: the main facts, at least one number or date where the topic has one, and any point where your sources disagree with each other.

Leave out: your own opinion, introductions that describe what you are about to say, and anything you did not actually find in the sources you read.

Return the brief in this format, every time:

— A one-sentence summary of the topic
— Three to five bullet points of key facts
— A line listing what you checked
— If your sources disagreed, one line saying how

If you could not find enough to answer properly, say so plainly rather than filling the format with weak material.

Why each part is there

"Writing for someone who needs to understand a subject quickly and will act on it." Giving the agent a reader changes the register of everything it produces. Without it you get prose that is technically responsive and useless to skim.

A number on how many sources. This is the line beginners leave out, and it is the one that causes the most trouble. "Research thoroughly" has no stopping condition, so the agent keeps reaching — which costs you against the run limit below and can stop the run before it finishes. Three is a sensible starting point. Watch your own usage and adjust; there is no universally correct number and we are not going to invent one.

An explicit "leave out" list. Models pad by default. Naming the padding is more effective than asking for concision, because "be concise" is a preference and "no introductions describing what you are about to say" is an instruction.

"Where your sources disagree." The most valuable line in the whole thing. A summary that flattens disagreement into confident consensus is actively misleading, and this is the one thing an agent will not do unless asked.

A format stated as a list. Specifying shape is what turns output you have to read into output you can scan. It also makes two runs comparable, which is what lets you tell whether an instruction change actually improved anything.

Permission to fail. Without the last line, an agent that found little will still fill the format. You want the empty result, because a confidently-formatted brief built on nothing is worse than no brief.

The one limit to know

On Zapier's free plan a single run can use up to 10"activities" before it stops and asks you for input rather than finishing, with 400 a month in total. An activity is the agent reaching outside itself — a search, a page read, an action. Talking to it is not counted.

That is why the source cap in the instructions is doing real work rather than being tidiness. A task with no stopping condition can reach the ceiling before it reaches the end of the job, and the pause that follows reads like a crash when it is not.

We have deliberately not told you what one research run costs in activities, because we have not measured it and will not guess. It depends on how many pages the agent chooses to read, which depends on how tightly you wrote the instructions above. The full free-tier breakdown covers what is actually known.

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.

Adapting it

Three changes worth making once the base version works, in order of how much they teach you.

Change the reader. Swap the first line for who actually receives this — a client, your team, yourself in six months. The whole output moves.

Change the format block. Ask for a verdict line first, or two bullets instead of five, or a "what I would check next" line at the end. Small format changes, re-run, compare.

Constrain the sources. Add a line about what to prefer or avoid — recency, source type, whatever matters for your subject. Then check whether the brief actually changed, because sometimes it will not, and finding that out is useful.

If you get stuck on the tool step

Everything above is typing. The part that is not typing is connecting the search tool, and it is where most first builds stall — the full walkthrough covers that block properly, and the beginner's orientation to Zapier Agents covers what to click if the interface is not matching what you expected.