How do you build an AI agent without writing code?

An AI agent is a program that uses a language model to decide what to do next and then does it in other tools: reading a spreadsheet, drafting an email, creating a task. A chatbot answers you; an agent acts for you. Building one used to mean writing code to connect the model to each app. No-code tools now handle those connections, so the work that remains is the part that really matters: defining the job clearly and deciding how much freedom the agent gets.

The best first agents are small, frequent, low-risk chores. Think "draft a reply to every new sign-up," not "manage my inbox."

The four parts of any agent

Whatever tool you use, you will set up the same four things.

1. A trigger. What starts the agent: a new row in a sheet, a new email, a message in a chat channel, or you asking it something.
2. Instructions. A plain-language description of the job, written like a brief for a new assistant: the goal, the steps, the tone, and what to do when unsure.
3. Knowledge and data. The documents or live data it may read, such as a price list, a FAQ or a spreadsheet.
4. Actions and permissions. What it may do in which apps, and which parts it may fill in on its own.

A worked example in Zapier Central

Zapier Central is an experimental workspace, in public preview, where you set up AI assistants that work across the thousands of apps Zapier connects to. Imagine a volunteer running sign-ups for a community workshop through a Google Form that feeds a Google Sheet.

Step 1. Create the assistant. Sign in at central.zapier.com with a Zapier account and create a new assistant called "Workshop sign-ups."

Step 2. Connect its data. Add the Google Sheet of sign-ups and a Google Doc with the workshop details (date, address, what to bring) as data sources, so the assistant answers from your facts rather than inventing them.

Step 3. Write the instructions.

You help run sign-ups for a free beginners' workshop on using AI tools.
When a new person signs up:
1. Read their name, email and the question they asked, if any.
2. Using only the workshop details document, draft a short, friendly confirmation
   email: thank them, repeat the date, time and address, and answer their question.
3. If their question is not covered by the document, do not guess. Write
   "I'll check and get back to you" and flag the row for me.
4. Keep every email under 120 words. Sign it "The workshop team".

Step 4. Add a behavior with a trigger and an action. The trigger is "new row in Google Sheets." The action is "create a draft in Gmail," not "send email." Zapier lets you control each field of an action: the assistant can always use a fixed value, choose from a set of values, generate the value itself, or leave it empty. Here, fix the "To" field to the email column from the new row, and let the AI write only the subject and body.

Step 5. Test. Add three test rows to the sheet: one with no question, one with a question the document answers, and one it can't answer. Read every draft. Check that dates and addresses match the document exactly, and that the third case was flagged rather than guessed.

Step 6. Supervise, then widen. Run it in draft mode for a week or two. Only when the drafts consistently need no editing should you consider letting it send, and even then keep the flag-for-me rule.

The same idea as a custom GPT

If the job happens inside a conversation rather than on a trigger, a custom GPT in ChatGPT may fit better. Paid ChatGPT users can build one at chatgpt.com/create by chatting with the GPT Builder, then refine it in the Configure tab: instructions, up to 20 knowledge files, and capabilities such as web browsing, image generation and data analysis. To let a GPT act in other systems, you add custom actions, which connect it to an outside service's API, a defined way for programs to talk to each other. Setting up an action means pasting a technical description of that API, so it is the one part that may need help from someone technical, or a ready-made option such as the Zapier AI Actions GPT, which OpenAI offered as an example when GPTs launched.

Limits and supervision

Three facts should shape how you use any agent. First, its output varies: Zapier itself notes that AI assistants can answer the same request differently each time, so one good test run proves little. Second, it can only be as accurate as the data you connect; OpenAI also warns that content from a GPT's knowledge files may appear in its answers, so never upload anything you would not want a user to see. Third, free tiers have ceilings. Zapier counts each run as an activity, with up to 400 a month on its free plan and 1,500 on Premium, plus preview limits such as five connected data sources.

Keep a person in charge of anything irreversible: sending, deleting, paying or publishing. Drafts, labels and flags are safe places to start.

Which no-code tools can run agents?

Two mainstream options are ready to try today. Zapier Central is built for agents that run on their own: assistants that react to triggers, read live data from tools such as Google Sheets, Google Docs and Notion, and act in the apps you have connected. It is still a preview, so expect changes.

Custom GPTs in ChatGPT suit conversational agents, where a person asks and the GPT researches, analyzes or acts during the chat. They are available to ChatGPT Plus, Team and Enterprise users, and can be shared privately, inside a company or publicly.

Other automation platforms are adding AI steps as well, and the four parts described above apply to all of them. When you compare tools, ask the same questions each time: what can trigger it, what data can it read, what can it do without asking, and how easily can you see what it did.