Build your first agent
Seven steps to a small, safe agent that takes work off your plate. No coding, with your approval.
Beginnerabout 1 hourAs of October 2026
What is an agent?
A chat answers a question. An agent works through a process: it gets a trigger (such as a new email), reads information, decides according to your rules and prepares an action. Good first agents are small, recurring and easy to check.
Your first agent in 7 steps
1. Pick a small task
Choose something that happens often, always runs in a similar way and does little harm if it goes wrong. Good: pre-sorting enquiries, summarising notes, suggesting appointments. Not for a start: payments, contracts, HR decisions.
2. Write down the process
Write down what happens today: trigger, steps, result. For example: new enquiry in the contact form → identify the request → draft a fitting reply → hand it to me for approval.
3. Choose the tool
For flows between apps, automation platforms with agent features fit well; for chat assistants, the agent features of the big AI providers. Look at connections to your apps, billing and privacy. Our comparison of agent builders helps you choose.
4. Give clear instructions
Describe role, goal, rules and limits in simple sentences. What must the agent never do? When should it ask? You can use our skills as ready-made instructions.
5. Keep permissions tight
Connect only the apps and folders that are really needed, and start with read access. Use a separate test inbox or test data if you can. Never put passwords in instructions; use the tool's sign-in instead.
6. Build in human approval
Let the agent create drafts instead of sending, deleting or paying on its own. Add an approval step before every outward action, such as a notification with “Approve” or “Reject”.
7. Test, start small, watch
Test with ten real but non-critical cases. Check the tool's log, set a cost or run limit, and expand only once the results are right. Note how to switch the agent off.
Example: pre-sorting contact form enquiries
- Trigger
- New message via the contact form
- Task
- Identify the request (quote, support, other) and draft a reply
- Instructions
- The skill “Reply to a customer message” with your facts on prices and delivery times
- Permissions
- Read the form inbox, create drafts, send nothing
- Approval
- You check every draft and send it yourself
- Success
- Replies in minutes instead of hours, no false promises
Safety check before you start
- Permissions: only the apps needed, read access first, no admin account.
- Data: no health, bank or ID data and no special categories of personal data unless strictly necessary. Check the provider's privacy notice and data processing terms.
- Approval: a human approves everything that goes out or costs money.
- Outside instructions: emails, web pages and documents can contain hidden instructions. The agent should treat them as material, not follow them.
- Costs: set a limit for runs or credits and understand how the tool bills.
- Control: review the log regularly and know how to stop the agent at once.
Legal: privacy and the AI Act
If your agent processes personal data, you are responsible for it, for example for a legal basis under the GDPR, a data processing agreement with the provider and informing the people concerned. If an agent talks directly to customers, tell them openly that they are communicating with an AI; the EU AI Act sets out transparency obligations for this. This guide is general information, not legal advice.