How to build an AI agent without writing code
You can build a working AI agent without code in three ways: draw it in a visual workflow builder, describe it in plain-language instructions, or show it by doing the task once while a tool such as Ditto AI records you. All three produce the same four parts: a model, instructions, tools and permissions. What differs is how much of it you have to spell out yourself.
What an AI agent is
An AI agent is a model that works toward a goal by choosing and calling tools, one step at a time, instead of only answering a question. A chatbot tells you how to file an invoice. An agent opens the email, saves the PDF to the right folder and adds a row to the spreadsheet.
Every agent, however it’s built, comes down to four parts:
| Part | What it is | Example |
|---|---|---|
| Model | The AI that decides the next step | Claude, GPT, or another model with tool use |
| Instructions | The goal and the procedure | “File each new invoice in Drive and log it in the expenses sheet” |
| Tools | The actions it can take | Read mail, upload a file, add a spreadsheet row |
| Permissions | What it’s allowed to do with those tools | Read the inbox; write only to one folder and one sheet |
“Without code” means you never write the glue between these parts yourself. You still decide what the agent does and what it may touch.
Three ways to build one without code
1. Draw it: visual workflow builders
Tools such as n8n, Zapier and Make let you connect boxes on a canvas: a trigger (a new email), a few actions (save the attachment, add a row), and, more and more often, an AI step that decides or writes something.
Good for: fixed flows between web apps that all have integrations, running in the cloud on a trigger.
The catch: you design every branch. If invoices sometimes arrive as links instead of attachments, that’s another branch you have to anticipate and draw.
2. Describe it: plain-language instructions
Instruction-based builders, such as custom GPTs in ChatGPT, or Projects and Skills in Claude, take a written description of the job plus a set of connected tools. The model follows your instructions and picks tools as it goes.
Good for: tasks you can explain precisely, where the tools are already connected.
The catch: writing down what you do by habit is harder than it sounds. The steps you skip in the description are the ones the agent gets wrong.
3. Show it: build by demonstration
With a demonstration builder, you do the task once while it records, and it writes the instructions for you. Ditto AI works this way on a Mac: it reads the recording, lists the steps and the inputs that change each time, finds a tool for each step, and checks that each tool works before you run anything.
Good for: repeated work across two or three apps that you already know how to do, but have never written down.
The catch: it can only learn what it saw. A recording of one idle window gives it nothing to repeat, and it needs a Mac with Apple silicon.
Which one to choose
| Draw it | Describe it | Show it | |
|---|---|---|---|
| What you do | Connect boxes on a canvas | Write the procedure | Do the task once |
| Where it runs | Usually the builder’s cloud | The chat app | Your Mac |
| Handles variations | Only the branches you drew | If your instructions cover them | The model adapts, within the permissions it was given |
| Best first task | “When X arrives, do Y” | A task you can explain in a paragraph | A task you do every week across apps |
How to build your first agent by showing it
This walk-through uses Ditto AI. The same thinking applies to any builder.
- Pick one task you did this week and will do next week. Filing an invoice, sending a weekly report, updating a tracker.
- Record it once, with real data. Start recording, do the whole task at your normal pace, and stop. If you talk while you work, say which values change: “this amount comes from the PDF”.
- Read what it understood. Ditto shows the process, each step with the frames it came from, and the inputs it found. Fix anything that’s wrong now.
- Connect the tools. For each app that needs your account, allow the tool server and sign in or paste a key. Each tool is checked with one read-only call and shows configured when it works.
- Run it with new inputs. Type or say what’s different this time, or leave it empty to repeat the recording. Turn on Ask before it creates, sends or changes anything for the first few runs.
- Watch the first run. Every step shows as it happens, and Stop ends the run along with every tool server it started.
For the detail behind each stage, see How Ditto AI works.
How to pick a good first task
A good first task is:
- Repeated. Weekly or more. A one-off task isn’t worth automating.
- Spread across two or three apps. That’s where copying and pasting eats your time, and where an agent helps most.
- Rule-based. You could explain it to a new colleague in a few sentences.
- Low risk if it goes wrong. Drafts, files and spreadsheet rows are easy to undo. Payments and deletions are not.
Examples that fit: filing an invoice from email into a Drive folder and logging it in a sheet; posting a short list of the day’s top stories to a Slack channel; turning a request into a Linear issue and keeping its link in Apple Notes.
Check these before any agent runs
Whichever way you build it, look at these before the first run:
- Which tools can write? Every write should be one you expect: this folder, this sheet, this channel.
- Where do credentials live? Keys should sit in your system keychain or the builder’s secret store, not in a note or a prompt.
- Can you stop it? You should be able to stop a run and know that everything it started has stopped too.
- Can you see what it did? Each step and each call should be visible afterwards.
The reasoning behind these is in Is it safe to let an AI agent use your accounts?
Common mistakes
- Automating a task you do differently every time. Pick the version you do most often and record that.
- Recording two tasks at once. One recording, one process. Record the second task separately.
- Using fake data. Test values hide the real variations, such as the invoice that arrives as a link.
- Letting it write on its own too early. Start with every write shown as a draft, then relax it once you trust the result.
Questions
Do I need to know how to code?
No. With a demonstration builder you do the task; with an instruction builder you write in plain language. Either way, you should be able to read what the agent will do before it does it.
Which model should I use?
One that’s good at tool use and, for demonstration, at reading screenshots. Ditto AI reads recordings with your own Anthropic, OpenAI or Amazon Bedrock account, or with Ditto AI’s.
Can I change the agent later?
Yes. In Ditto you can edit the steps, inputs and tools before any run, or record the task again if it has changed.