How do I build an AI agent that does real work for me, without writing code?

One paragraph of plain English replaced the code Nick wrote no HTML and no Python. He typed one block of normal English describing the job, and Copilot came back with a proposed system, a capability list, and the apps it would connect. The build in the video runs across 9 steps, and none of them involve a line of code. The hardest typing was naming the spreadsheet columns he wanted.

The agent picked its own schedule When the builder needed a time of day, it asked itself the question and picked Monday at 9:00 a.m. without being told. Nick had named the cadence as weekly, not the hour. That is the part he found striking: Copilot and the Zapier agent system traded details with no human in between.

The free plan stopped the test at 10 activities Mid-run, the agent hit a cap of 10 activities per run on the free agents plan. A paid agents plan allows up to 40 continuous activities per run. The test still surfaced five names, filled rows in real time, and put a finished draft in Gmail.

What are the steps to build the agent?

These are the 9 steps as Nick lays them out in the video, in order, with the timestamp where each one starts. Copy the order. The snags show up in the same places for most people.

All 9 steps from the video, kept in order.

Where are the key moments in the video?

Ten timestamps, in order, if you would rather watch the parts that matter than the whole build.

All 10 key moments from the video.

Can you build an AI agent without coding?

Yes. Nick built a working research and outreach agent in Zapier Copilot by typing a single paragraph of plain English describing the job. He wrote no HTML and no Python, and he says outright that he does not know how to write either.

What the builder did on its own is the part to notice. Copilot wrote the workflow instructions, set the trigger, and wired up Google Sheets and Gmail. It also chose to use web search to find experts and to run the job weekly. Those choices showed up in a section called assumptions made, so Nick could read them before he agreed.

The human work that was left was small, but it was real. He had to connect the accounts, create the Google Sheet himself, add the column headers, point the Sheets action at the right file, and approve each action during the test run. None of that is coding. All of it is the kind of clicking you already do in a browser.

So the honest answer is not that the agent builds itself. The answer is that the typing you do is normal English, and the system turns that into steps.

This build was sponsored by Zapier, and Nick says so in the video. You can follow along at the sponsor link below and build the same thing in your own account.

What are AI agents, and what can they actually do?

An agent differs from an automation in that it makes decisions rather than following a fixed path. A plain automation moves a piece of data from one app to another. An agent looks at the world, judges what it finds, and then acts.

The one in this video searches the web on its own, judges whether a person is a good fit for the show, writes what it finds into a spreadsheet, and composes a different email for each prospect. Nick's framing: it is not a macro, it is closer to a junior researcher who never gets bored of the same task.

Here is the concrete job. Every Monday the agent hunts for people who could appear as guests on an AI podcast. It looks for a large platform, a book in the space, a book about to come out, or past podcast appearances about AI. Then it fills a row per person and writes an invitation.

On the first test it found Melanie Mitchell, Lex Fridman, Andrew Ng, Yann LeCun, and Max Tegmark, filled the rows in real time, and put a finished draft in Gmail. Those are real names with real public work, not filler.

The shape of the job is what carries over. Any task that means searching, judging, recording, and then writing a first draft fits this same pattern. The agent is a pattern, not a podcast tool.

How do Zapier AI agents work?

You give the agent a goal in plain English; it proposes a set of capabilities, picks the apps and actions it needs, and asks you for anything missing. That back-and-forth is the whole system.

In this build it chose a Monday 9:00 a.m. schedule without being asked, requested details about a sample prospect so it could personalize the email template, and flagged when it could not reach a spreadsheet. You approve each action before it runs.

The flow has two halves that talk to each other. Copilot takes your paragraph and drafts a plan. Then you click continue with agent builder, and the plan is handed to the agent system, which writes the workflow instructions and builds the steps. Nick's read on it: Copilot and the Zapier agent system are talking to each other and working out the details with no human in between.

The trigger it picked was a Schedule by Zapier weekly trigger. That is the clock that wakes the agent up. Everything after the trigger is the agent deciding what to search, what to keep, and what to write.

What you keep is the veto. During the test run it asked for approval before each action, so nothing happened that Nick had not seen first. That is the setting to keep on while you are learning what your agent does.

How do I write a prompt for an AI agent?

Describe the outcome the way you would explain it to a new hire. Nick's prompt named the task, the selection criteria, every spreadsheet column he wanted, how many results per run, how often to run, and where the output should land.

Look at step 2 in the table for the full shape. He asked for an agent that researches potential guest experts on artificial intelligence, people with large platforms, a book in the space, or podcast appearances about AI. Then he listed the columns by name: first name, last name, date discovered, follower count, social accounts, contact email, a summary of accomplishments and why they fit the show.

Last came the cadence and the output. Find five potential guests each Monday, and draft an email to each one, saved under drafts in Gmail. That one sentence set the schedule, the volume, and the finish line.

Being that specific is what let Copilot propose a complete system on the first pass. When he realized he had forgotten a criterion, he simply added another sentence and it updated the plan. The missing piece was people with books about to come out, and the builder folded it into the capabilities with no restart.

Do not try to write instructions like code. Brackets, pseudo-code, and all-caps commands do not help here. One block of normal English does.

How do I create my first AI agent, step by step?

The video lays the build out in 9 steps, and the table above holds all of them in order. If you want the short version, the 9 steps group into five moves, and the five moves are easy to hold in your head.

First, write the job out in plain language, including what you do not want it to do. Second, let the builder draft the agent's capabilities and correct anything it got wrong. Nick added a requirement mid-build and the builder absorbed it.

Third, connect the apps it needs, which here were Google Sheets and Gmail. Nick's Gmail and Google Drive were already linked to Zapier, so the agent could reach both. Linking an app is just logging into it through the Zapier interface and granting access.

Fourth, test it before it runs live, because it will get something wrong; in this build it could not find the spreadsheet and had to be pointed at it manually. Fifth, turn it on, which is a separate action and easy to forget.

Read step 6 and step 7 in the table before you start, since that is where the one real snag lives. The middle of the build has more small clicks than the summary suggests. Working at your own pace will take a little longer than watching.

How do I automate Google Sheets with an AI agent?

The agent writes a new row per result. In the video it fills in each prospect's name, background, recent work and contact details as it finds them, and you can watch the rows appear in real time during the test run.

One gotcha to know: the agent could only add rows to a sheet that already existed, so Nick had to create the sheet and its column headers first. He asked it to create a fresh spreadsheet so the data would not land in an existing one, and it could not do that.

So he made the sheet himself, named it for the project, and typed in the column headers the agent had listed. That is a small piece of manual work, and you have to do it before the run. Skip it and the agent has nowhere to put anything.

The payoff is a living record. Every Monday the sheet grows by a handful of rows, each one with the same columns in the same order. You can sort it, filter it, and share it like any other spreadsheet.

Nick also accepted the agent's own suggestion to add a step that checks the existing sheet before each run so it does not pull duplicate names. That one extra step keeps the list clean over months of runs.

Why can't my Zapier agent find my Google Sheet?

This happened to Nick on camera. The agent cannot create a new spreadsheet, it can only add rows to one that already exists, and after he created his sheet the agent still could not see it.

The agent was honest about it. It told Nick the core research and outreach functionality was ready and offered to save the agent and fix the sheet afterward. That is a reasonable offer, but do not leave it broken and walk away.

The fix was manual: open the agent settings, open the Google Sheets create spreadsheet row action, search for the sheet by name directly, and select it. He saved the agent without testing until that was corrected. After selecting the file, he let the agent pick the worksheet value itself.

Searching by name is the key move. The file list inside the action did not surface his new sheet on its own, so the search box is the way in. Once the sheet was selected, the rest of the action filled in.

Then test again. During the run, open the spreadsheet in another tab and check the spreadsheet while it runs to confirm rows are actually landing. Seeing a row appear is the proof that the connection took.

How do I automate Gmail with an AI agent?

It writes to drafts, not to send. The agent composes a personalized invitation for each prospect using what it learned about them and saves it in Gmail drafts, where a person reviews and sends it.

The build sets that up in step 5 of the table. The agent saves finished emails to Gmail drafts instead of sending them, and the video calls that choice deliberate. Nothing leaves your account on its own.

On the test run it put a finished draft in Gmail. Nick opened it and read it, which is the moment at 1102 seconds in the key moments table. The draft drew on what the agent had found about the person.

Nick is emphatic about this: never let an agent send on your behalf. The agent does the research and the typing; you decide what actually goes out. The goal is to remove the drudgery, not the judgment.

The pattern holds for any outreach you dread. The agent finds the person, learns who they are, and writes the first version. You still read each draft and press send yourself.

Should I let an AI agent send emails for me?

Nick says no, and he built his agent to stop at the draft. That choice is step 5 in the table, and the video calls it deliberate.

His words: he does not plan to ever have AI send emails on his behalf, no matter how well it can write in his voice, because he always wants to make edits and keep the human element. The agent removes the research and the first draft. You still review, personalize, and hit send.

The draft folder is where that edit happens. The agent saves finished emails to Gmail drafts instead of sending them, so nothing leaves your account until you say so.

The time math still favors you. The slow parts of outreach are finding the person, learning who they are, and staring at a blank message. The agent does all three. Reading and tweaking a ready draft takes a minute.

Set this rule before you build, not after. If the prompt says save under drafts in Gmail, the agent builds it that way from the start, and you never have to go back and remove a send step.

Is the Zapier agents free plan enough to test an agent?

It was enough for Nick to build and start testing, but he hit a ceiling mid-run. The message said he had reached the limit of 10 activities per run on the agents free plan, and that a paid agents plan allows up to 40 continuous activities per run.

An activity is one action the agent takes, such as a search, a row written, or a draft created. A job that researches five people and writes five emails moves through those fast. That is why the cap showed up during the first real test rather than during the build.

His take was that if the agent does what he hopes, the price pays for itself. That is a sensible way to decide. Build on the free plan, watch one honest test run, then judge whether the output earns the monthly cost.

The key moments table puts this at 762 seconds, so you can watch the exact message appear. Knowing what the wall looks like keeps you from thinking your agent is broken when it simply stopped.

Plan your first test around the cap. A smaller ask lets you see the whole path from search to draft inside the free limit.

How much do you have to supervise an AI agent?

More than zero, and Nick is direct about the frame. Treat it as an intern who needs managing, not an executive who handles everything. That line is the heart of the video.

Count the human touches in this build. He had to connect accounts, create the sheet, add the headers, fix the Sheets selection, and approve each action during the test, and he still plans to check every guest the agent finds before reaching out.

The point is how little intervention it took, not none. Most of those touches happen once, during the build. After the agent is on, the weekly run is the agent's job and the review is yours.

Keep the approval prompts on while you are new. During the test the agent asked for approval before each action, which lets you see what it wants to do before it does it. Once you trust the pattern, you can loosen that.

Supervision also means reading the output. Nick plans to check every guest the agent finds before he reaches out to anyone. The sheet is a starting list, and the last read is still yours.

How do I turn the agent on when the build is done?

A finished agent is not a running agent. Nick had to click the enable button to switch it on, and only then noticed he had never named it, so he named it after the job it does.

This is step 9 in the table and it is the easiest step to miss. The build screen looks complete, the test worked, and nothing tells you the schedule is still asleep. Click enable, then check that the toggle actually changed.

Name the agent for the job, not for the app. Your agents list shows everything you have built, so a name drawn from the job tells you what a given agent is for at a glance.

Then add the cleanup step. Nick accepted the agent's own suggestion to add a step that checks the existing sheet before each run so it does not pull duplicate names. That keeps repeat names out of the list.

After that, the work is the weekly review. On Monday the sheet has new rows and Gmail has new drafts. You read, edit, and send. That is the whole point of building the thing.

Method and sources: The source is a single build captured on video by Nick Standlea, published in November 2025 and reviewed here as of October 2026. The population is one agent built in Zapier Copilot for one job: weekly research on potential podcast guests, rows written to Google Sheets, and invitation emails saved to Gmail drafts. The step rows and the timestamps are copied from the video in the order they occur, and the key moments table keeps all 10 markers. This page excludes pricing beyond the free and paid activity caps shown on screen, because the video does not state a dollar amount; the exact cost of a paid agents plan is not reported here, and neither is the time the build took on its own. The build was sponsored by Zapier, and results on other accounts may differ depending on which apps are connected.