How do AI agents work?

An AI agent works in a loop: it reads a goal, decides the next step, uses a tool (your calendar, inbox, website), looks at the result, and repeats until the job is done or it needs to ask you. The language model supplies the judgment; the connected tools supply the hands.

The four parts of every AI agent

Strip away the marketing and every AI agent has the same four parts. A model that can read and write language. A set of tools it is allowed to use. A memory of what has happened so far. And a loop that keeps it going until the goal is met.

1. 1GoalYou give it an outcome, not a command. "Fill the two empty slots on Thursday" rather than "send this text to these people."
2. 2PlanThe model breaks the goal into steps: check who cancelled recently, check who asked for Thursdays, draft a message, send, watch for replies.
3. 3ActIt calls a tool for each step. Read the calendar. Search the customer list. Send a WhatsApp. Each tool is a small, permissioned door into your business.
4. 4CheckIt reads the result of each action and decides what to do next. If a customer replies "can you do 3pm instead?" it adjusts rather than stopping.

A worked example: filling a cancelled appointment

A physio clinic gets a cancellation for 2pm tomorrow. A rule-based automation could text a waitlist. An AI agent does something closer to what a good receptionist does. It checks who on the waitlist asked for afternoons, drafts three personal messages, sends them, and when the first person says yes, books them, updates the calendar, and tells the others the slot has gone. Total owner involvement: zero, unless the agent is set to ask before sending.

How the model "decides"

The model is a large language model, the same family of technology behind ChatGPT, Gemini and Claude. Given the goal, the tools available and everything that has happened so far, it predicts the most sensible next step. It is not consulting a rulebook; it is pattern-matching against a vast amount of text about how tasks like this get done. That is why it copes with the unexpected and why it occasionally needs a human to say "not that."

What tools can an agent use?

• Read and write to a calendar or booking system.
• Read and send email, WhatsApp or SMS.
• Look up customers, prices and hours in a database or CRM.
• Edit a website page or publish a social post.
• Search the web or read a document you have given it.

Each tool is a permission. A well-designed agent platform lets you decide which tools it can use freely, which need approval, and which it can never touch.

Where the loop stops

Good agents stop in three situations: the goal is met, they hit something outside their permissions, or they are genuinely unsure. That third one matters. An agent that never asks is a liability; an agent that asks about everything is just a slower version of you. The right setting is usually "act on routine, ask on money, promises and anything public."

Agents vs. automation, in one sentence

Automation runs a fixed recipe. An agent reads the situation and writes the recipe as it goes. Automation is cheaper and more predictable for tasks that never vary; agents earn their keep on tasks that are similar every time but never identical, which describes most of the admin in a small business.

Where Kahu fitsKahu AssistKahu Assist is an AI agent that runs your website, marketing, bookings and follow-ups from one chat, and asks before anything goes live. You keep the yes; it does the typing.Connects to your site, WhatsApp, calendar and emailAsks before anything goes liveSet up from one 15-minute chatSee Kahu Assist →Fill Thursday’s empty slots and let the cafe know the new brunch hours.Kahu · 3 actions readyTwo Thursday gaps at 11:00 and 14:30. Drafted messages to 6 waitlisted customers.Updated the hours on your site and Google profile to Sat–Sun 8–2.One post about the new hours, using your Sunday counter photo.ChangeApprove all

Frequently asked questions