How to build an AI agent with no code? And where to start?

You can build an AI agent with no code in an evening. But first, honestly: do you need one, what to automate first, and where to build it (Cowork, Codex, n8n).

Can you build an AI agent in one evening, without a single line of code?

The short answer: yes. The long answer: if you start straight from building, you'll most likely build the wrong thing, in the wrong place, for the wrong task. Because in this topic 80% of the result isn't "how to build" but "what and why". I'll cover the "how" in detail too, but let's do it honestly and in order.

One thing up front, so you don't waste your time. If you search "how to build an AI agent with no code", you'll get a dozen articles that, under the guise of a tutorial, sell you a specific chatbot builder. I won't do that: I don't have my own builder to push. I have experience building agents for real tasks, and that's what I'm sharing.

And one more thing: if you're not yet sure what an agent is and how it differs from a chat and a project, I have a separate article about that with criteria. I won't repeat it here, straight to the point.

The very short version

Building an agent with no code is real: for that there's Claude Cowork, OpenAI's Codex and n8n (more on them below). But building is the last step, not the first. First, three questions:

Do you even need an agent?
What part of your work hurts the most?
What of that can be automated right now?

Go through them and you'll build something useful. Skip them and you'll build a toy that runs for a week and dies. Start with the head, not the tool.

Think first, build later

This is the most boring and the most important part. Picking a tool is easy, but picking the right task is half the battle. Let me break it into three steps.

Step 1. Do you even need an agent

Sounds odd in an article about building an agent, but it's the honest first question. An agent isn't needed for everything, and that's fine. A quick fork:

The task is solved in one message ("rewrite this text", "explain this"), that's an ordinary chat, no agent needed.
You need a helper with your context on one topic, that's a project (or a Custom GPT).
An event always triggers the same prescribed steps ("a form submission goes into the CRM"), that's an automation (n8n, Zapier), often with no AI at all.
But when the steps aren't known in advance and the model has to decide for itself what to do next, that's where you need an agent.

If in doubt, I covered the criteria for an agent separately. Don't chase the word "agent" just because it's trendy. Sometimes the smartest solution is a formula in a spreadsheet.

Step 2. Find what hurts the most

Don't automate what's "trendy to automate". Automate what actually eats your time and annoys you.

A simple 15-minute exercise. Write down everything you do by hand and regularly over a working week. Next to each item put two marks:

how much time per week it eats;
how much you dislike it (on a scale from "tolerable" to "hate it").

The top of that list (lots of time + very annoying) is your candidates. Not an abstract "adopt AI", but a concrete "every Friday I spend two hours assembling a report from exports by hand".

Step 3. What of that can be automated now

Not everything at the top of the list can be handed to an agent today. Run the candidates through a filter:

Repeatability. Are the steps roughly the same, with only the data changing? Good.
Predictable result. Is it clear what "done well" looks like? Good.
Access to data. Will the agent have somewhere to pull information from (a spreadsheet, email, CRM)? If the data is scattered across your head and sticky notes, gather it first.
Cost of a mistake. If a mistake is expensive (money, legal actions), that doesn't mean "you can't", it means "always with a human's confirmation".

What passes the filter is your first task. One. Not five. One working process from start to finish beats five half-built ones.

And how exactly to frame that task, I'll give you a ready-made prompt at the end of the article: it will question you itself and put together a short brief, which you can either build yourself or send to me.

Where to build with no code

Now about the tools. "No code" doesn't mean "a terminal for programmers". For an ordinary person there are three real doors, from the simplest to the more flexible. Codex and n8n also have more technical modes, but you can start in a human way.

Tool	What it is	Barrier to entry	For whom
Claude Cowork	A desktop app with an agentic loop and access to your local files	Low, it's an interface	Working with files on your computer: reports, documents, spreadsheets
Codex (OpenAI)	OpenAI's agent: works with your files and processes, supports MCP and the open Skills standard, assembles repeatable processes	Medium, a bit more technical	Those on the OpenAI stack who are ready to go a little deeper
n8n (the AI Agent node)	A visual process builder: you connect a model and tools, and the agent chooses what to call	Medium, but no code	Wiring services together and running on a schedule or by event

A separate note on chatbot builders (BotMan, Athena and the like), which search will show you first. They aren't bad, but they solve a different task: a chatbot on a website or in a messenger that answers clients. That's not the same as an agent that does the work for you. If you need a bot for your site, go there. If you need a helper that assembles reports and sorts email, that's Cowork, Codex or n8n. Choose by the task, not by the ad.

How to build: one example, step by step

Let's work through a live example. Take the task "every Friday, assemble a weekly report from an export in a spreadsheet". It passed all the filters: it repeats, the result is clear, the data exists, a mistake isn't critical. Let's go step by step (screenshots of my real build are below).

Set up a workspace and context. A folder (in Cowork) or a project: put a past report, a description of the metrics, and formatting rules in there. The agent needs to see what "good" looks like.
Write the instruction. In your own words: what we take as input, what we calculate, in what form we return it, what to avoid. This is the body of the task.
Package the repeatable part into a skill. If the method is the same every week, assemble it once as a skill playbook, and the agent will pull it up itself. No need to explain again.
Connect data and services through MCP. For the agent to reach your spreadsheet, email or CRM, you need connectors. How this works with no code, I covered in the article about MCP.
Test on real data. Give it last week and compare with what you did by hand. Doesn't match? Fix the instruction, don't settle.
Put a confirmation button on the risk. Anything to do with money, campaigns and important emails, the agent shows you and waits for a "yes". This isn't paranoia, it's about your safety.

Advice from Anthropic that I stand behind: start simple. Don't build a ten-task monster right away. Assemble one working process, polish it, and then repeat the same method for the next one.

How much it costs and how long it takes

Honest orders of magnitude, without "we'll deploy it in 3 days and you'll strike gold":

Time for the first build: from a couple of hours to an evening, if the task is chosen right and the data is at hand.
Money: the no-code part you'll build on your own paid subscription (Cowork, or ChatGPT on the Pro plan and up). If it's n8n with integrations, add an inexpensive server and pay-as-you-go API, usually tens of dollars a month, not thousands.
Where the money grows: complex integrations, rare scenarios, a high cost of mistakes. Then either dig in deeper, or call in a person.

A prompt that will question you itself

If you're still thinking about where to start, don't keep it in your head. Copy the prompt below, paste it into Claude or ChatGPT, and answer the questions. You'll get a ready brief for an agent as the output.

Help me choose and describe one task for an AI agent. Ask one question at a time, and at the end put together a short brief.

1. Which routine task eats the most of my time and annoys me the most?
2. What do I give as input (a file, a link, an email, an export)?
3. What result do I want as output?
4. What can the agent do on its own, and where must it stop and ask me (money, campaigns, important emails)?
5. How will I know it's done well?
6. What must the agent never do?

At the end: state the task in one paragraph and suggest which tool is easiest to build it with (Cowork, Codex, n8n), or whether it's better handed to a person.

From there, two paths. Either you take that brief and build it yourself by the steps above. Or, if you'd rather not bother, you send it to me and I build the agent. The brief helps in both cases: you now know exactly what you need.

When you don't need an agent

So you don't waste an evening for nothing:

The task is one-off, one request in a chat will do.
The steps are strictly predictable and identical, that's automation (n8n, Zapier), AI is often unnecessary there.
Accuracy matters more than flexibility, and the data is simple, a script or a spreadsheet formula is more reliable.

An agent is justified where the route can't be guessed in advance and someone has to think on the fly. In all other cases, take a simpler tool, it's cheaper and more predictable.

What it looks like in practice for me

So I'm not just theorizing: I covered my own working example separately. It's my morning AI agent, which every morning sorts through my task list, tidies up the wording and suggests a plan for the day. There you can see exactly this method in action: one task, context, tools and a confirmation button where needed.

FAQ

Do I need to know how to program?

No. Cowork is an app where you write in plain words. n8n is a visual builder, you connect blocks with the mouse. Codex is closer to development, but it's also driven in natural language. Code is only needed for your own complex tools; for a first task it isn't required.

How much does it cost to start?

If you already have a paid subscription to Claude or ChatGPT (Pro and up), you'll do the first build with no extra spend. For n8n integrations, add an inexpensive server and pay-as-you-go API usage.

Are a chatbot and an AI agent the same thing?

No. A chatbot answers your clients in a dialog. An agent does the work for you: it gathers, calculates, formats, acts. The builders from search are more often about the first. Decide what you need before choosing anything.

Where do I start today?

Not with a tool. Write down your weekly routine, mark what takes the longest and annoys you the most, pick one task and run it through the ready prompt above. The tool will fit itself to the task in five minutes.

Is it safe to give an agent access?

It's safe if you give access at a minimum (only for the task) and put a confirmation before actions with money and campaigns. Treat it like hiring a new employee: access as trust grows, important decisions through you.

If you've read this and realized the task is there, but you'd rather not bother with the build, I build agents like this end to end. Tell me what the routine is, and I'll suggest a solution: order agent development or just write to me on Telegram. No "you urgently need an agent", just honestly, for your task.