Back to blog

AI agents in simple words: what they are and how to start

Everyone keeps saying "AI agents". I wanted to understand what that actually means, so I tried a few. Here is what I found, without the jargon.

The chore that got me curious

I own a house with a few tenants, and every month I did the same electricity bill routine. I took photos of each submeter. I read the total from the MSEB bill. I split it by hand, worked out one tenant's share on the MSEB website, and then typed a separate WhatsApp message to every tenant.

It was not hard. It was just boring, and boring jobs are exactly what AI tools are good at. That is what pushed me to look properly at AI agents, and then to build a small billing app with AI helping me along the way.

So what is an AI agent?

You already know chatbots. You ask a question, it gives an answer, and then it waits. If you want something done, you do it yourself.

An agent is different. You give it a goal, and it goes and works on it. It can open a website, edit a file, run a program, or check its own work. When it is finished, it comes back and tells you what happened.

ChatbotYou ask a question→Chatbot answers→You do the work
AgentYou give a goal→Agent uses tools→Reports back, job done
A chatbot tells you how. An agent does it.

How an agent actually works

Under the hood, an agent is just a loop. It looks at the goal, picks the next small step, does that step with a tool, and then checks what happened. If the job is not finished, it goes around again.

That is the whole idea. It is also why agents can feel so different from chatbots. They are not answering once. They keep trying, fixing, and checking until they think it is done.

Your goal→Pick next step→Use a tool→Check result→Not done? Go again. Done? Tell you.
The loop: pick a step, do it, check it, repeat.

Three kinds you can try today

1. Coding agents. Tools like Claude Code, Cursor and GitHub Copilot's agent mode can read a whole project, change several files, and run tests. Code is a good place for agents because the computer tells them straight away if something is broken. Of the three kinds, these are the most reliable right now.

2. Browser and computer agents. ChatGPT's agent mode and Claude's computer use can click around websites, fill in forms and collect information. They are handy for repetitive online tasks, but they are slower and make more mistakes than coding agents.

3. No-code automation tools. Apps like n8n and Make let you connect an AI model to your email, spreadsheets or calendar by dragging blocks around. You do not need to write code, and they suit small jobs you repeat often.

What changed for my billing chore

I built a small web app for my monthly bills. Now I type in the submeter readings and the bill amounts, and it splits the cost, adds any old balance, and writes the message for each tenant. I copy each one into WhatsApp and I am done.

Before
  1. Photograph each submeter
  2. Read the total from the bill
  3. Split the bill by hand
  4. Check the MSEB calculator
  5. Type every message myself

Five steps, lots of typing

After
  1. Photograph each submeter
  2. Enter readings and amounts
  3. Copy the ready messages

Three steps, the app does the maths

Same chore, fewer steps. The boring parts moved to the app.

I did not need to be an expert to do this. I described what I wanted in plain words, looked at what came back, and asked for changes. Some of the first versions did not work, and I had to say so and ask again. That back and forth is normal.

How to start, step by step

  1. Pick one small task. Choose something boring that you do again and again. Avoid anything where a mistake would cost you money or upset someone.
  2. Say what "done" looks like. Tell it what you want, what to use, and what to leave alone. "Tidy my spreadsheet" is vague. "Remove duplicate rows and sort by date, but do not delete anything else" is much better.
  3. Watch the first few runs. Do not just read the final answer. Look at what it actually did, so you learn where you can trust it.
  4. Make the next task a little bigger. Once one task works well, try a slightly harder one.

Stay safe while you experiment

An agent can take actions, so it can also take the wrong ones. A few habits help a lot.

  • Give it only the access it needs. If it only has to read one folder, do not give it your whole drive. For my billing app I use a GitHub token that works on a single private repository and nothing else.
  • Approve anything you cannot undo. Sending emails, making payments and deleting files should always need your OK.
  • Be careful with what it reads. A web page or document can hide instructions meant to trick an agent. Treat what it reads as information, not as orders.
  • Keep backups. If something goes wrong, you want an easy way back to how things were.
Agents are helpful, but they still make mistakes. Always check the important results yourself, especially anything to do with money.

Where I think this is going

The tools are getting better fast, and there are now shared standards that make it easier to connect an agent to the apps and data you use. I do not think agents will replace careful people. I think people who know how to work with them will get more done, and get bored less often.

If you have been waiting on the sidelines, my advice is simple. Pick one small, dull task this week and let an agent try it. You will learn more from that one experiment than from reading ten more articles, including this one.