From one task to a working AI agent: the Trigger, Data, Action, Deliver pattern
An AI agent is a repeated task in four parts: a Trigger, the Data, the Action, and the Deliver. Learn the pattern and build your first Claude agent, no code.
An AI agent is just a task you repeat, broken into four parts: a Trigger that starts it, the Data it reads, the Action it takes, and where it Delivers the result. To turn a manual task into a Claude agent, write those four lines, then build the smallest version that runs once from start to finish.
Every simple agent, whether it drafts replies to reviews or builds your morning brief, runs on the same shape. Let's walk through it on a real example and build a first agent you are not afraid to switch on.
What is the Trigger, Data, Action, Deliver pattern?
The four parts that make up any simple agent:
- Trigger - what starts it. A schedule (every morning at 8), an event (a review lands, an email arrives), or your manual "go".
- Data - what it reads to do the job. Your inbox, a sheet, a product page, a document with your tone of voice.
- Action - what it does with the data. Judges, sorts, drafts, calculates, translates.
- Deliver - where the result goes. A draft to approve, a chat message, a row in a sheet, a notification.
That is all a simple agent is. If you can name those four things for your task, you have already designed the agent.
How do I turn one of my tasks into an agent?
1. Pick a task you do by hand every week. Narrow and boring: sorting reviews, compiling a report, answering routine emails. 2. Write four lines. Trigger, Data, Action, Deliver, one phrase each. If you cannot phrase one, the task is not ready to automate yet. 3. Build the smallest version. One review, one run, a start you press by hand. Not the whole process at once. 4. Make Deliver a draft, not a final action. The agent prepares, you send and confirm. 5. Run it on real data and fix what breaks. Then add volume and remove the manual start.
What does it look like on an example?
Take an agent that replies to new reviews.
| Part | What it is in the example |
|---|---|
| Trigger | a new review appears (or once a day the agent checks for new ones) |
| Data | the review text, the product page, a doc with your reply tone |
| Action | read the sentiment and draft a reply in your voice |
| Deliver | a draft reply you check before it goes live |
Start with drafts, and switch on auto-reply only once you trust how the agent writes.
What makes a good first agent?
Reversible and low-stakes. A good first agent prepares a draft instead of taking an irreversible action.
- Reversible: a draft, a label, a row in a sheet. Not sending money, deleting, or emailing a client unchecked.
- Boring and frequent: the thing that eats time every week, so the payoff shows at once.
- With a clear Deliver: you can say in one sentence what you get out of it.
Where does it usually break?
- Too big a task at once. Build the path for one case, then widen it.
- Deliver straight to production. For the first weeks let the agent hand you a draft and you decide.
- A vague Action. "Deal with the reviews" means nothing to an agent. "Judge the sentiment and draft a reply under 300 characters in a friendly tone" does.
- No data on hand. If the agent has nowhere to read your tone or product page, the reply comes out generic.
In short
An agent is a four-line task: Trigger, Data, Action, Deliver. Write those lines for one of your routines, build the smallest version with a draft as output, and run it on real data. When it works, add volume and drop the manual start.
You can build one right inside Claude, no code: connectors give it the Data, skills and assistants give it the Action. AGINE Academy covers this on your own tasks in the real Claude, and the starting block is open without signing up, where you can test the pattern on a live example.
Questions
A prompt is one request you type by hand each time. An agent is the same prompt with a Trigger that starts it on its own and a Deliver that sends the result where it belongs, so it runs on a schedule or an event and carries the job to the finish.
No. A simple agent is built right inside Claude with no code: connectors give it access to data, and skills and assistants set the action. Coding comes in when the agent grows into a separate product, but the first working version you can build in plain words.
A reversible, boring one you do every week. A good first agent prepares a draft instead of doing something irreversible: it writes the reply, you send it. That keeps the cost of error near zero while it saves time right away.