How to build an AI agent in 7 steps (with or without code)

To build an AI agent you need four things: a clear task, a model, functions and tests. Building it is not the hard part: making it work well with real customers is. This guide follows the order that avoids rework.

Step 1. Pick a single task

The most common mistake is wanting an agent that does everything. Start with a repetitive, measurable, low-risk task:

• Answer frequent questions and book appointments.
• Qualify leads that arrive on WhatsApp.
• Give order status.

Done = you can describe the task in one sentence and say how you know it went well.

Step 2. Choose how to build it

Not sure? Read Should you build or hire an AI agent?.

Step 3. Define its functions

List the actions it needs and nothing else. For a booking agent: check_availability, book_appointment, cancel_appointment, hand_off_to_human. Each with a clear description and parameters with a format. See Functions in AI agents.

Step 4. Write its instructions

The instructions (the system prompt) say who it is, what it does, what it doesn't do and when it hands off to a person. Write them as if you were training a new employee:

• Its goal in one sentence.
• Which data it must have before acting.
• What it must never do.
• When to ask a person for help.

Step 5. Give it memory carefully

Memory lets it remember the customer, but it also piles up junk: repeated data, contradictions or sensitive information it should not keep. Decide what it stores and review it regularly.

Step 6. Test it with cases it has never seen

This is the step almost everyone skips. Testing with your own messages doesn't work: the agent passes the test you already know it will pass. You need:

1. Real or realistic cases the agent has not seen.
2. An objective way to grade: did it call the right function? with the right data? did it ask for what was missing?
3. A number: the share of correct answers before and after each change.

Without a number you don't know if a change made the agent better or worse. See Best tools to test AI agents.

Step 7. Launch with limits and measure

Launch at a low autonomy level (acts with approval or acts and reports) and move up only with data. Log every action. See How much freedom does an AI agent have.

Measure your agent before launching

AUN measures how your agent calls its functions on a hidden test, returns a fix that keeps names and parameters unchanged, and measures again. You pay per call and only if the report exists.

• Audit my agent
• Review my agent's memory
• Already running? List it and sell it on AUN