Before I get into the how-to, let me set your expectations correctly.
"AI agents" is one of those terms that's getting stretched to cover everything from a two-step Zapier workflow to a fully autonomous system running complex multi-step tasks in the background. What we're building today is real and genuinely useful — but it's not the sci-fi version you see in the YouTube thumbnails.
An AI agent (in the no-code world, anyway) is an automated workflow that uses an AI model to make decisions. Instead of hard-coded rules that say "if subject line contains 'invoice,' forward to accounting," an AI agent reads the email, understands the context, and figures out where it should go. That's the difference. The AI adds judgment.
OK. Let's build one.
What Is an AI Agent, Actually?
An agent has three core components:
A trigger — something that starts the workflow. An incoming email. A new row in a spreadsheet. A message in Slack. A timer that fires every hour. Without a trigger, nothing runs.
A brain — an AI model (usually GPT-4o, Claude, or Gemini) that reads the input and decides what to do. This is what makes it an "agent" rather than a regular automation.
Actions — things the agent does after the brain decides. Send an email. Update a database row. Post a message. Create a task. Call an API. The actions are how the agent affects the world.
That's it. Trigger → Brain → Actions. Every agent you'll ever build, no matter how complex, is some version of this.
The "no-code" part means you're connecting these pieces using a visual drag-and-drop interface instead of writing Python. The logic is the same. The syntax just looks like building blocks instead of text.
Which No-Code Platforms Actually Work?
There are a lot of options. Most of them aren't worth your time as a beginner. Here's my honest take on the four platforms that matter:
Make.com (my pick for beginners)
Make.com is where I'd send every non-developer who wants to build agents. The visual canvas is the best in the industry — you can literally see data flowing between modules as you test. It has 1,900+ app integrations, direct OpenAI and Claude connectors, and a free tier that gives you 1,000 operations/month.
The learning curve is maybe 2-3 hours before you're building real things. After a week, you'll be comfortable with most of what it can do.
Start building on Make.com for free →
Zapier AI
Zapier is the most beginner-friendly of the bunch — if something's easy to explain in a sentence, Zapier can probably do it. Their AI features (Zapier AI Actions, the AI by Zapier step) are solid for simple classification and routing tasks. But the free tier is stingy (100 tasks/month), and complex logic gets unwieldy fast.
Good for: Simple one-step AI tasks, people already using Zapier for basic automations.
Not great for: Multi-step agents with branching logic.
n8n
n8n is the power tool. It's open-source, self-hostable, and has the most flexibility of any platform here. If you're a developer or someone comfortable with technical concepts, n8n is where you eventually want to land. The AI nodes are excellent, and the ability to run it on your own server means no per-task pricing.
The catch: steeper learning curve, more configuration required, less hand-holding for beginners.
Relevance AI
Relevance AI is built specifically for AI agents, not general automation. Their interface is more like "describe your agent in English and it builds the workflow" than "drag modules onto a canvas." It's impressive for creating agents that do research, write content, or handle customer questions.
Worth trying if your use case is AI-heavy (lots of reasoning, long prompts, complex outputs). Less useful if you need deep integrations with other software.
Let's Build an Agent in Make.com: Step by Step
We're going to build an email triage agent. It watches your inbox, reads each new email, and automatically:
- Labels it (Support / Sales / Personal / Spam)
- Drafts a reply if it needs one
- Creates a task in your project management tool for anything that requires action
This is a legitimately useful agent. I use a version of this myself. And it's buildable in about 2 hours on your first attempt.
Step 1: Create a Make.com Account
Go to make.com and sign up. The free tier is more than enough to build and test this agent.
Once you're in, click Create a new scenario. That's Make's word for a workflow.
Step 2: Set Up the Gmail Trigger
- Click the large + in the center of the canvas
- Search for "Gmail" and select it
- Choose Watch Emails as the trigger
- Connect your Gmail account (Make will request permission — this is standard OAuth, your password never leaves Google)
- Set Folder to "Inbox"
- Set Maximum number of results to 5 (start small while testing)
You'll see a sample email appear in the trigger panel. This is Make showing you what data is available from the trigger — the email's subject, body, sender, timestamp, etc. Make a mental note of what's here. The AI step will use all of it.
Step 3: Add the AI Classification Step
- Click the + after the Gmail trigger
- Search for "OpenAI" (or "Claude" if you have an Anthropic key — both work)
- Choose Create a Completion (OpenAI) or Create a Message (Claude)
- Connect your API key (you'll need an OpenAI or Anthropic account — the free tiers are fine for testing)
Now write your prompt. This is the "brain" step. Here's what I use:
You are an email classifier. Given the email below, output a JSON object with exactly these fields:
- category: one of "support", "sales", "personal", "newsletter", "spam"
- priority: one of "high", "medium", "low"
- needs_reply: true or false
- summary: one sentence describing what this email is about
Email subject:
Email sender: ()
Email body:
Output only valid JSON. No explanation.
The ,, and are Make's variable syntax. When you click inside the prompt field, you can pick these from the Gmail module's data. Just start typing — click the variable picker to find it)
- Click Generate data structure and paste in a sample of what your AI returns
After this step, Make knows that category, priority, needs_reply, and summary are separate data points it can use in later steps.
Step 5: Apply a Gmail Label
- Add a Gmail → Add a Label module
- Set Message ID to the Gmail trigger's message ID variable
- Set Label to `` from the parsed JSON
Make sure the labels exist in Gmail first. Go create "support," "sales," "personal," "newsletter," and "spam" labels in Gmail before running the scenario.
Step 6: Draft a Reply (Conditional)
This step should only run when needs_reply is true. Otherwise you'll draft pointless replies to newsletters.
- Add a Router module (the diamond icon) after the label step
- Create a route with the condition: `
equalstrue` - On that route, add another OpenAI/Claude module
- Write a prompt that drafts a reply:
You are a professional assistant drafting email replies. Write a brief, helpful reply to this email.
Original email from :
Subject:
Body:
Write a reply that is:
- Professional but warm
- Under 150 words
- Does not make any specific commitments
- Ends with a clear next step
Output only the reply text, no subject line.
- After the AI step, add a Gmail → Create a Draft module
- Set To to the sender's email, Subject to
Re:, Body to the AI reply
Step 7: Test It
Click Run once at the bottom of the screen. Make will check your inbox for new emails and process them.
Watch each module light up as it runs. Click on any module to see the exact data that passed through it. When something breaks (and it will, at least once), this is how you debug.
If everything works, you'll see your emails labeled and drafts waiting in Gmail.
Step 8: Activate the Scenario
Once you're happy with the test, click the toggle at the bottom left to turn the scenario on. Make will run it on a schedule (every 15 minutes on the free tier, down to every minute on paid).
Your agent is live.
Common Mistakes (And How to Avoid Them)
Not handling errors. Real emails are weird. Your AI prompt will occasionally return malformed JSON. The email body will sometimes be HTML instead of plain text. Add Make's built-in error handling (right-click on a module → "Add error handler") to any step that might fail. Without this, one bad email breaks the whole scenario.
Writing prompts that are too vague. "Classify this email" is not a prompt — it's a wish. Give the AI exact categories, exact output format, and examples when possible. The more specific you are, the more consistently it performs.
Building the complex version first. Everyone wants to build the 15-step agent immediately. Start with the 3-step version. Get it working. Add one step at a time. The 15-step agent you actually finish is worth more than the ambitious one you abandon.
Forgetting about cost. OpenAI and Anthropic charge per token. An email triage agent processing 50 emails a day costs almost nothing (fractions of a cent per email at gpt-4o-mini rates). But if you're feeding long documents to expensive models at high volume, audit your costs before you scale.
Not testing with real data. Make's built-in test data is clean and predictable. Real emails are not. Run your agent against at least 20 real emails before considering it ready for production.
What AI Agents Can and Can't Do Right Now
They're good at: Classification and routing. Summarization. Drafting text from templates. Converting unstructured inputs (emails, PDFs) into structured data. Monitoring and alerting. Consistent execution of repeatable tasks.
They're not good at: Tasks requiring deep contextual memory across many interactions. Complex multi-step reasoning where each step depends heavily on the last. Creative work that benefits from iteration and feedback. Anything requiring physical actions in the world.
The wall you'll hit: No-code AI agents break down when the task requires genuine judgment — not pattern matching, but actual decision-making with ambiguous inputs and meaningful consequences. For those use cases, you still need a human in the loop.
That's not a knock on these tools. It's just an honest description of where the technology is in early 2026. The classification and routing use cases are genuinely solved. The autonomous decision-making stuff is still being figured out.
What to Build Next
Once your email triage agent is running, here are some agents worth building:
Content monitoring agent — Watches RSS feeds or Google Alerts for mentions of your brand or keywords, summarizes them daily, routes important mentions to Slack.
Lead enrichment agent — When a new lead comes in through your website form, looks them up against LinkedIn data, categorizes them by size/industry, and routes them to the right sales rep.
Meeting prep agent — 30 minutes before a calendar event, pulls in the attendee's LinkedIn profile, recent email threads, and company news. Drops a summary in Slack.
Document classification agent — Watches a Google Drive folder, reads incoming documents, tags them by type, and moves them to the right subfolder.
All of these are buildable in Make.com without writing a single line of code. The email triage agent you just built is the foundation — same three-part structure, just applied to different triggers and outputs.
For a deeper look at automation platforms overall, check out our best AI agent platforms roundup. And if you're ready to graduate to code-based agents, our guide to building full-stack apps with AI covers the developer path.
Make.com is an affiliate partner. If you sign up through our link, we earn a commission at no cost to you. It doesn't influence what I write — I'd recommend Make regardless.
Top comments (0)