Guide · made with Zapier
How do Zapier AI agents work? What Nick's build showed
How Zapier AI agents work, seen through Nick's build: plain-English instructions, a weekly trigger, Sheets and Gmail actions, and real limits.
The short answer
A Zapier AI agent is an assistant you describe in plain English, and Zapier turns that description into written instructions, a trigger that wakes it up, and a set of actions it is allowed to take in your connected apps. When it runs, it makes its own choices inside those limits: it searches the web, judges what it finds, writes results into apps like Google Sheets, and drafts things like emails. Nick saw all of that in one build. He typed a paragraph asking for an agent that finds five possible AI guests each Monday, logs them in a Google Sheet, and drafts invitations in Gmail. Zapier Copilot proposed the system, listed the assumptions it made, wrote the instructions, and set a weekly Schedule by Zapier trigger, picking Monday at 9:00 a.m. on its own. On the test run the agent searched, filled spreadsheet rows in real time, asked Nick to approve its actions, and put a draft in Gmail. It also showed its limits. It could only add rows to a sheet that already existed, it could not find the new sheet until Nick picked it by hand, and the free plan stopped the run at 10 activities. Nick's frame for the whole thing: "It's an intern, not an executive."
Build an agent with Zapier →Built with Zapier, the sponsor of this video.
Step by step
It reads a plain-English job and proposes a system ▶2:38
Nick typed one paragraph describing the job. Zapier with Copilot came back with a proposed system: an agent that researches AI experts weekly using web search, evaluates them against his criteria, writes Google Sheets entries, drafts personal outreach emails, and saves them in Gmail, on a weekly Monday schedule. It named the apps it would use, Google Sheets and Gmail, with web research built into the agent itself. Nothing has run yet at this point. It is a plan to read.
It tells you what it assumed ▶3:22
Under a heading called assumptions made, Copilot listed the choices it filled in by itself. It would use web search to find experts, run each Monday, find contact details through publicly available sources, and base each email on the person's background. Reading this, Nick noticed a gap. He had left out people with books about to come out, so he typed that in, and it was added to the agent's capabilities. The assumptions list is where a wrong guess can be caught before it becomes behavior.
It writes its own instructions and sets a trigger ▶4:19
After Nick clicked continue with agent builder, the system wrote the workflow instructions for him. Then it created the trigger, the thing that wakes the agent up: Schedule by Zapier, every week. The screen briefly said no triggers found, then it sorted that out with no help from Nick. Next it added the actions it would need, starting with Google Sheets. Zapier's own guide describes agents the same way: instructions, a trigger that starts the work, and actions the agent can carry out.
It answers its own questions when it can ▶4:54
Nick watched the builder reason through gaps. It noted that it needed a time of day for the Monday run, listed a few options, and then set a reasonable time itself: Monday at 9:00 a.m. Nick had not given it an hour. "It's basically talking to itself," he said. His read is that Copilot and the Zapier agent system pass details back and forth and work them out with no person in between.
It asks you when it cannot decide ▶6:06
Some gaps it could not fill alone. It asked for details about one prospect so it could personalize the email. It asked which existing spreadsheet to use, or whether Nick wanted a new one, since the current options did not seem right for the job. It asked what tone he preferred for podcast invitations. Zapier's guide describes the same habit in agents that are already running: a Needs action list shows tasks waiting on you, such as missing information or a connection that needs fixing.
It runs into the edges of its tools ▶7:39
Nick asked the agent to make a new sheet. It said it could not. It could only add rows to spreadsheets that already existed. So Nick made the sheet and added the column headers. The agent still could not find it. It finished the Gmail steps first, said it would come back for the sheet, and then offered to save the agent and fix the sheet later. Nick opened the create spreadsheet row action, searched for the sheet by name, and selected it. An agent can only do what its connected actions allow.
On a run, it searches, judges, writes, and checks in ▶10:57
Nick clicked preview, then test agent. The agent ran a Google search, dug through websites, and came back with Melanie Mitchell first, with her social accounts, a follower count, and a summary of why she fit the show. It asked Nick to approve before it ran the action, then moved on to the next names, Lex Fridman among them, while rows appeared in his spreadsheet in real time. Partway through it hit a wall: the free agents plan allows 10 activities per run, and a paid agents plan unlocks up to 40 continuous activities per run.
It drafts, and then you switch it on ▶18:01
For each prospect the agent ran the Gmail create draft action and saved an invitation in Nick's drafts. It sent nothing, because Nick set it up that way. Nick had told it nothing about his show, yet it found the show online and described it in the email, though it leaned on college preparation, a topic he has covered a few times. The agent then reminded him to turn it on if he was happy with its work. Nick raised one more point, checking the sheet before each run so no guest is picked twice, and the agent added that step. Then he clicked enable. A built agent does nothing until it is enabled.
“I've always been at the forefront of automation with my business to make things run more smoothly, but this is next level because this agent is actually making decisions.”
Nick Standlea, in the video
Key moments
- ▶1:32The one-paragraph job Nick typed
- ▶2:38Copilot proposes a system
- ▶3:22The assumptions it made on its own
- ▶5:24Copilot and the agent system talk to each other
- ▶6:06The agent asks Nick what it needs
- ▶7:39It can only add rows, not create a sheet
- ▶9:19It cannot find the new sheet
- ▶10:57The first test run starts
- ▶12:40Free plan stops at 10 activities per run
- ▶18:01Gmail create draft runs
- ▶18:58Nick on an agent that makes decisions
- ▶20:14A step added to avoid duplicate guests
An agent is instructions, a trigger, and actions Nick's paragraph became the agent's instructions. Schedule by Zapier, every week, became its trigger. Google Sheets and Gmail became its actions, with web search built in. Zapier's own guide describes every agent with the same parts, plus optional data sources the agent can read.
The builder filled some gaps on its own When it needed a time of day for the weekly run, the builder set Monday at 9:00 a.m. without being told. It also wrote down the assumptions it made, like using publicly available sources for contact details, so Nick could check them before going further.
It asked when the choice was Nick's The agent did not guess at everything. It asked for details about a sample prospect, which spreadsheet to use, and what tone Nick liked for invitations. Knowing which questions to hand back is part of how it works.
Its limits were the limits of its tools It could add rows to a Google Sheet but could not create one, and it could not find the sheet Nick made until he selected it by hand. An agent can only do what its connected actions allow.
It stops at the draft, by design Nick built the agent to save invitations in Gmail drafts and never send them. He reviews each guest and edits each email himself, and he plans to give the agent his own email template next.
The free plan capped the test at 10 activities Partway through the first test, the agent hit the limit of 10 activities per run on the agents free plan. A paid agents plan unlocks up to 40 continuous activities per run.
What is an AI agent in Zapier?
Zapier describes its agents as assistants you train with prompts to do specific jobs across its apps. Each one is built from a few parts: instructions in plain language, a trigger that starts it, actions it is allowed to take in connected apps, and, if you want, data sources and web browsing.
Nick's build made those parts easy to see. His paragraph became the instructions. Schedule by Zapier, every week, became the trigger. Google Sheets and Gmail became the actions. Web search came built in, so the agent could look for people on its own.
What sets an agent apart from a fixed automation is the judgment in the middle. After the test run, Nick described the agent looking at possible guests and saying no to some and yes to others. A plain automation moves data along a set path. This agent chose who belonged on the list.
Copilot sits on top of all this. It is Zapier's building assistant, and it turned Nick's paragraph into a proposed system, a list of assumptions, and then a working agent. Zapier's guide says Copilot can fully configure an agent in minutes from plain instructions.
The apps have to be connected first. Nick's Gmail and Google Drive were already linked, and he notes that linking an app is just logging into it through Zapier and granting access.
What happens when a Zapier agent runs?
The trigger fires first. In Nick's agent that is a weekly schedule, set for Monday at 9:00 a.m. Between runs, it waits.
Then the agent works through its instructions. On the test it ran a Google search, dug through websites, and built a profile for each candidate: social accounts, a follower count, a summary of the person's work, and why they would fit the show. Melanie Mitchell came first.
Each finding turns into an action. The agent wrote a row to the Google Sheet, and Nick watched the fields fill in while the agent moved on to the next name. Then it ran the Gmail create draft action and left an invitation in his drafts.
During the test, it checked in. It asked how each action looked and whether to change anything before it ran, and Nick approved. It also asked for the app permissions it needed.
Runs have a limit. On the agents free plan, Nick's test stopped when it reached 10 activities per run. A paid agents plan unlocks up to 40 continuous activities per run.
Where are Zapier agents weak?
The tools set the ceiling. The agent could not create a new spreadsheet, only add rows to one that existed. After Nick made the sheet, it still could not find it, and he had to select it by hand inside the create spreadsheet row action.
Its guesses can drift. With no information about the show, the agent found it online and wrote it up, but it framed the show around college preparation, a subject Nick has covered a few times. It got a follower count from Twitter only. The first email was long, and Nick plans to give it his own template next.
It does not avoid repeats unless told to. Nick raised the risk of picking the same guest twice, and the agent added a step to check the sheet before each new search.
Zapier's own guide is plain about the trade. It suggests a Zap when you need precision and an agent when about 80% accuracy is good enough and speed matters more than perfection.
So the person stays in the loop. Nick keeps every email as a draft, checks each suggested guest himself, and treats the agent as an intern who needs managing, not an executive who handles everything.
Method and sources: This page draws on one sponsored build video by Nick Standlea, published November 10, 2025, read from its full transcript, plus Zapier's Agents page and Agents guide for product facts; it covers one agent and one test run, so results in other accounts may differ.
Sources and references
Related guides
Questions people ask
How do Zapier AI agents work?
You describe a job in plain English, and Zapier turns it into instructions, a trigger, and actions in your connected apps. When the trigger fires, the agent works through the instructions and makes its own choices along the way. In Nick's build the trigger was a weekly Monday schedule. The agent searched the web for possible guests, judged who fit, wrote each one into a Google Sheet, and drafted an invitation in Gmail. It asked him when it could not decide, and it stopped at the draft.
What are AI agents and what can they actually do?
An AI agent takes a goal and works out the steps itself, instead of following a fixed path. Nick's agent shows the range. It searched the web, weighed candidates against his criteria of large platforms, books, and podcast appearances, filled spreadsheet rows, and wrote a personal email for each person. Zapier's template library points the same pattern at other jobs, such as researching new sales leads or drafting support replies from a knowledge base.
How do I build an AI agent without coding?
Type the job in plain English and let the builder do the rest. Nick wrote one paragraph in Zapier with Copilot and wrote no HTML and no Python. Copilot proposed the system, wrote the instructions, and set the trigger. He still had to connect his apps, create the Google Sheet, and fix one setting by hand. The companion guide linked below walks through his build in order.
What is the best no-code AI agent builder?
This page cannot rank builders, because it rests on one build in one tool, in a video Zapier sponsored. What it can show is what worked in Zapier for Nick. He typed one paragraph and got a working agent with no code. The builder used the Gmail and Google Sheets accounts already linked to his Zapier account, and it listed its own assumptions before building anything. The weak spots were real too: it could not create a spreadsheet and needed help finding one. If you are comparing builders, try each on a real job of your own and watch where it asks for help.
What does a Zapier agent need before it can run?
Connected apps come first. Nick's Gmail and Google Drive, which holds Google Sheets, were already linked, and linking an app is just logging into it through Zapier and granting access. It also needs instructions, which Copilot wrote from his paragraph, and a trigger, here a weekly schedule. Anything its actions depend on has to exist already, which is why Nick created the spreadsheet and its column headers himself. Last, the agent has to be switched on with the enable button.
What is the difference between Zapier Copilot and a Zapier agent?
Copilot is the building assistant, and the agent is what it builds. Nick typed his request to Copilot, which proposed a system and listed its assumptions. When he clicked continue with agent builder, the agent was set up with its own instructions, trigger, and actions. During the build, Nick described Copilot and the agent system as talking to each other and working out the details with no human stepping in.
Does a Zapier agent make its own decisions?
Within its instructions, yes. Nick's agent picked Monday at 9:00 a.m. for its schedule when he had not named an hour, and on the test run it chose which people to include. Nick said this is what set it apart from the automation he already used in his business: the agent was making decisions, passing on some people and keeping others. It still asked him when a choice was his to make, like which spreadsheet to use or what tone to take.
Where do Zapier agents fall short?
In Nick's build the limits came from its tools and its guesses. It could not create a spreadsheet, only add rows to one that existed, and it could not find his new sheet until he selected it by hand. The first email framed his show around college preparation, a subject he has covered a few times. The free plan stopped the test at 10 activities per run. Nick's answer is to keep checking its work.
Should a Zapier agent send emails on its own?
Nick says no. He set his agent to save every invitation in Gmail drafts, and he says he does not plan to ever let AI send email for him, no matter how well it writes in his voice. He wants to edit each one and keep the human element. Zapier's guide gives similar advice: ask the agent for approval before critical actions like sending emails.
What happens when a Zapier agent gets stuck?
In Nick's build, it said so and offered a way forward. When the agent could not find his new Google Sheet, it finished the Gmail steps first, said it would look again, and then offered to save the agent so the sheet could be fixed later. It told him how: open the Google Sheets action in the agent settings and change which sheet it uses. For agents that are already running, Zapier lists stuck tasks under Needs action, such as missing information or a connection that needs attention.
How many actions can a Zapier agent take in one run?
It depends on the plan. Nick's test hit the limit of 10 activities per run on the agents free plan. The message said a paid agents plan unlocks up to 40 continuous activities per run. Nick's job of researching guests, writing rows, and drafting emails reached the free limit partway through the first test. He said that if the agent does what he hopes, it will be worth the cost.
When should you use a Zapier agent instead of a Zap?
Zapier's guide puts it this way: use a Zap when you need precision and predictability, and an agent when about 80% accuracy is good enough and speed matters more than perfection. Nick's build sits on the agent side. Finding and judging possible guests is work that calls for judgment, and every result still passes through him before anything goes out.
Zapier sponsored this video. Nick chose what to build and what to say.