Define the job in one sentence
What it delivers and what it starts from. If it doesn't fit in one sentence, that's two agents.
Guide · Setting it up
How to create an AI agent: five decisions, zero code
Can you create an AI agent without coding? Yes. It comes down to five decisions, put in writing: what it's for, when it steps in, how it works, what it can't do, and how you test it. If you can explain a job to a new hire, you already have the hard part.
The example in this guide is a made-up neighborhood bookstore.
At a glance
How to create an AI agent for my business: the five steps
Say when it should step in
The signal that tells it this one's for it: a kind of message, a kind of request.
Explain how it works
The steps and the format, in the order a person would do them.
Set the limits
What it can't promise or decide without asking you.
Test it with hard cases
Two normal ones and one of the kind that ruins your afternoon. Before you show it to anyone.
What an agent is, and what you need first
An AI agent is an assistant with a fixed job and its own instructions: answering reviews, preparing quotes, checking copy before it goes out. What sets it apart from a plain assistant is that you don't have to explain the job every time. The full comparison is in AI agent vs chatbot vs assistant.
One note on the word: in the US, “AI agent” often suggests a bot that chats with customers. The agents in this guide work inside the team. They draft; a person reads and sends.
Before you create any, one thing has to be true: the assistant knows your company. A special-order agent that hasn't read your suppliers' lead times will write, very fluently, a made-up lead time. Documents first; then the agent.
The example in this guide is Paper Lantern Books, a neighborhood bookstore with a book club and special orders.
Decision 1: the job
One sentence that says what it delivers and what it starts from. “Helps with books” isn't a job. “Drafts the reply to someone asking for a title we don't have in stock, with the lead time and the deposit” is.
If you need an “and” to join two different jobs, that's two agents.
Decision 2: when it steps in
The signal that tells it this one's for it. At the bookstore: someone pastes a customer's message asking about a book that isn't in the store. The more specific the signal, the less often it wanders in where it shouldn't.
Decision 3: how it works
The steps and the format, in the order a person would do them: where it looks, what it writes, how it signs off. This is where you can tell who knows the job. “Reply in a friendly way” doesn't help. “Look up the lead time in the suppliers document and write two paragraphs, the second one with the deposit” helps a lot.
Decision 4: what it can't do
Set the limits you'd give someone in their first week. Don't promise a date the distributor hasn't given. Don't offer discounts that aren't written down. Don't treat an order as placed without a deposit. An agent with no limits ends up promising something in your name.
There's one more limit, and it depends on the tool. Some agents carry out actions in other programs: they send the email, they create the order. Others prepare the work and leave it ready for a person to review. For a first agent, the second kind is the careful choice.
Decision 5: how you test it
With three cases before you show it to anyone: two normal ones and a hard one. At the bookstore, the hard one is the customer who ordered ten days ago and asks, annoyed, where their book is. If it invents a date there, go back to decision 4.
The key idea: creating an AI agent is a briefing problem, not a technical one. When it gets stuck, it's almost never the tool: it's that nobody had written down how that job gets done.
No-code AI agent builders: three families
We don't give a ranked list of brands, because it changes every quarter. What doesn't change is that there are three ways to create agents, and it helps to know which one you're in:
| Family | How you create it | Who it's for | What's on you |
|---|---|---|---|
| Form-based builders | You fill in fields: name, instructions, files | People who know the job and have patience to fine-tune | Writing the instructions well |
| Tools for developers | With code or workflows, wiring the agent to other programs | Technical teams that want it to carry out actions | Building it and maintaining it |
| Creating by talking | You explain the job and answer what it asks you | People who know the job and don't want to configure anything | Reading the proposal before you approve it |
How to create an AI agent with ChatGPT, Claude, Copilot or n8n
People usually search for this with a brand name attached. Here's where the four most common ones fit, per their own documentation (checked on November 12, 2026). The five decisions apply to all of them.
- ChatGPT. The form-based route has been the custom GPT. OpenAI says it's transitioning custom GPTs to “plugins,” so read its migration guide before you build on them.
- Claude. A Project is a workspace with its own instructions and knowledge files: first family. Anthropic describes it in its help center.
- Microsoft Copilot Studio. Microsoft calls it a graphical, low-code studio, and says you can create an agent by describing it in plain language. See the Copilot Studio overview.
- n8n. You build a workflow around an AI Agent node and connect a chat model and at least one tool to it: second family, and you maintain it. See the n8n documentation.
Flipa is in the third family: you ask for the agent in the chat, it asks you one to three questions with options to click, and it shows you a proposal to approve. Its agents prepare the text; sending it is your call. We explain it on the agents page, and there's a full case in Flipa for schools. Worth knowing before you ask: the app's interface is in Spanish today (English isn't available yet).
Common mistakes
- One agent for everything. “Customer support” is six jobs. Start with the one that repeats the most.
- Instructions that are wishes. “Be professional and helpful” changes nothing. Steps do.
- Testing only the easy case. The hard one is the one that shows up on a Friday afternoon.
- Never looking at it again. When a supplier or a term changes, change the profile.
When it isn't worth creating an agent
If the job comes up twice a year, ask the assistant when the time comes. If it's different every time, there's no fixed job to write down. And if all you want is a request one click away, saving the request is enough.
Checklist before you call it done
- The job fits in one sentence.
- It's written down when it steps in and when it doesn't.
- The steps name documents that exist.
- There are at least two limits in writing.
- It has passed two normal cases and a hard one.
- Someone on the team knows this agent is theirs to look after.
What it looks like in writing
Anatomy of an agent: the profile, part by part
This is the profile of the bookstore's special-order agent, with the five parts an agent has in Flipa. Open each one to see what it's for, and switch to the weak version to see the difference.
Agent profile · Paper Lantern Books
Special orders
NameSpecial orders
What it's for. So you can recognize it. Under each answer you can read which agent replied; if the name says nothing, neither does that notice.
Weak“Agent 3”
Good“Special orders”
What it doesDrafts the reply to someone asking for a title we don't have in the store, with the lead time and the deposit.
What it's for. It's the job in one sentence. If it doesn't fit in one, that's two agents.
Weak“Helps with books.”
Good“Drafts the reply to someone asking for a title we don't have in the store, with the lead time and the deposit.”
When it steps in on its ownWhen someone pastes a customer's message asking for a book that isn't in the store.
What it's for. It decides when it's the agent's turn without you calling it. Too broad, and it wanders in where it shouldn't. Too vague, and it never steps in.
Weak“When needed.”
Good“When someone pastes a customer's message asking for a book that isn't in the store.”
How it worksLooks up the distributor's lead time in the suppliers document. Writes the reply with that lead time and the deposit. If the distributor gives no date, it doesn't promise one: it says we'll be in touch.
What it's for. The steps, the format and the limits. What it can't do is written here too.
Weak“Reply friendly and fast.”
Good“Looks up the distributor's lead time in the suppliers document. Writes the reply with that lead time and the deposit. If the distributor gives no date, it doesn't promise one: it says we'll be in touch.”
Sample questionsCan you get me a mushroom field guide I saw at a fair? · How long does an imported book take? · I ordered it ten days ago, any news?
What it's for. They're for testing it and for showing the team what to ask it. Include at least one hard one.
Weak“Do you have books?”
Good“Can you get me a mushroom field guide I saw at a fair? · How long does an imported book take? · I ordered it ten days ago, any news?”
With this profile, anyone on the team knows what the agent does, when it steps in and what it won't promise.
And by talking?
The same decisions, answering questions
This is a mockup, not the product: you answer three questions and the profile of a bookstore agent fills in as you go.
What job are you giving it?
How do you want it to write?
What shouldn't it do without asking you?
Agent profile
Special orders
- How it talks
- Warm, like a bookseller making a recommendation
- Never without asking you
- Promise a date without checking the distributor's lead time
- What it already knows
- Everything you've told Flipa about your company
Which one first
Four first agents that make sense
Flipa has templates to get you going. They're starting points: they adapt to your company with what Flipa already knows about it, asking two things at most. Only the review one has been tested from start to finish; the others haven't been fully tested yet.
FAQ
Before you create your first one
Is it hard to create an AI agent?
Technically, no: there are tools where you don't write a single line of code. The hard part is the usual one: being clear about the job, the steps and the limits. If you can explain that job to a new hire, you can create the agent.
Where can I build an AI agent without coding?
In form-based builders, where you fill in a name, instructions and files, and in tools where you create agents by talking. In Flipa it happens in the chat: you ask for it, answer one to three questions and approve the proposal. The app's interface is in Spanish for now.
Can I create a personal AI agent for my own work?
Yes. The five decisions apply just the same to an agent only you use, say one that prepares your meeting notes. In Flipa, agents belong to the company's account and today they're created by whoever administers it; opening that up to the rest of the team exists as an option, off by default.
How long does it take to create an AI agent?
Writing down the five decisions takes half an hour if you know the job. In Flipa the interview is one to three questions, and having the assistant test the agent itself takes half a minute. What takes the most time is watching it work on real cases for a few days.
How do I know my agent works before I use it?
Give it two normal cases and a hard one, and read all three answers in full. In Flipa you can chat with the draft yourself without saving it. Or have Flipa test it: it runs three requests, one of them hard, and tells you for each whether it's fine and why.
If you can explain the job to a new hire, you can create an agent
Write the first decision today: the job, in one sentence.
Start free