How Do AI Agents Work? The 5-Step Loop Behind Every Task, Explained Simply

You typed "find me a cheaper energy deal than this bill" and something strange happened. A browser window opened by itself, a comparison site filled in, a tariff table appeared, and a minute later there was a short note on your screen with three options and a recommendation. It felt like magic, which is a problem, because things that feel like magic are hard to trust and easy to misuse.

This guide explains how AI agents work in plain English, with no computing background assumed. You will see the five-step loop that every agent runs, what the "model" inside it actually does, why it needs tools such as a browser, how its memory works, why it sometimes makes things up, what the credits you are charged pay for, and where you fit in. The sources are the companies that build the agents and the regulator that measures how Britain uses them, all read on 8 October 2026.

How we checked this guide

1. Builders' own descriptions. Anthropic's published guide to building agents for how the loop is designed.
2. Research on mistakes. OpenAI's September 2025 paper on why language models state false things with confidence.
3. Usage figures from Ofcom. The regulator's Online Nation report, published in December 2025.
4. Our own runs. Every example below is a task we have given an agent and watched.

The 5-step loop

Anthropic describes agents as systems where the model directs its own process and its own use of tools, working in a cycle of acting, getting feedback from the environment and acting again, with checkpoints for a human. In everyday language, that is five steps. Take the energy bill as the example throughout.

Step 1: it reads the goal and decides what "done" means

The agent turns your sentence into a target it can check against: "a tariff for this postcode and this usage that costs less than the current one, with the switching conditions listed". If the goal is vague, this step is where things go wrong, because the agent will guess at the target.

Step 2: it makes a plan

A short list of steps: read the bill for the usage figures, open a comparison site, enter them, collect the results, check exit fees, write up the three best. Good agents show you this plan before they start. Jobbit has a plan mode that stops here for your approval.

Step 3: it uses tools

This is the part that separates an agent from a chatbot. The agent can only talk, so to do anything it calls a tool: a web browser to visit the comparison site, a file reader to look at the bill, a calculator or spreadsheet to compare annual costs. Each tool returns a result, and the result goes back into the agent's thinking.

Step 4: it checks

After each step the agent compares what it got with what it expected. The comparison site asked for a meter type it does not know, so it goes back to the bill. One tariff is cheaper per unit but has a higher standing charge, so it recalculates over a year. This checking is why agents take minutes rather than seconds, and it is where most of their value comes from.

Step 5: it reports and waits

It writes the summary, shows its working if asked, and stops at the point where money changes hands or a message is sent. Switching the tariff is your decision. A well-built agent asks; a badly built one assumes.

Then the loop repeats for the next goal. Our plain-English guide to what an AI agent is puts this loop beside twelve real examples.

The engine: what the model actually does

Inside every agent is a model, the thing that reads and writes language. Claude, GPT and Gemini are models. A model has read an enormous amount of text and learned to predict what comes next, one small piece at a time. Those pieces are called tokens, and a token is roughly three quarters of an English word. Everything the agent "thinks" is written out in tokens: the plan, the reasoning, the tool requests, the final answer.

Three consequences follow, and they explain most of what beginners find puzzling.

• The model knows nothing after its training date unless it looks. That is why the agent opens a browser rather than answering from memory, and why answers about last week need a search.
• The model has a context window, the amount of text it can hold in mind at once. Think of it as a desk. Your bill, the plan, the web pages it has read and its own notes all sit on the desk, and when the desk is full the oldest papers fall off. Long tasks are managed by tidying the desk, which agents call compaction. Jobbit shows a context meter so you can see how full the desk is.
• The model is confident by design. It produces the most likely next tokens, and "most likely" is not the same as "true". The section on mistakes below explains what this means for you.

Ofcom's Online Nation report shows how ordinary this engine has become: about 30% of UK searches now show an AI overview, and 53% of adults say they see those summaries often. The model that writes the summary is the same kind of model that drives an agent.

Tools: how an agent reaches the world

Every row has a "what can go wrong" because tools are where an agent meets the untidy real world. The browser row deserves one extra note: a growing number of websites block agents by default, as our report on the web locking AI agents out explains, which is why the agent sometimes reports that it could not get in. Jobbit's agent browser lets you watch it work and take the controls yourself when a site needs a human, described in our guide to computer-use AI agents.

Memory: what it remembers and what it forgets

Beginners expect an agent to remember them the way a colleague would. Mostly it does not, and knowing the three kinds of memory saves a lot of repetition.

• Working memory is the context window from the section above. It lasts for one task and is wiped when the task ends, unless the agent writes notes to itself.
• Saved instructions are files you give it that it reads at the start of every task: who you are, how you like things written, what it must never do. On Jobbit these live in Projects, together with shared files, notes and any passwords it needs stored as secrets rather than typed into chat.
• Long-term memory is a feature some products offer, where the agent keeps notes across conversations. It is useful and worth checking occasionally, because a wrong note persists.

The practical rule: anything you find yourself typing twice belongs in saved instructions.

Why it sometimes makes things up

The confident wrong answer, called a hallucination, is the agent's best-known flaw. OpenAI's research paper on the subject, published in September 2025, gives the clearest explanation: models are trained and tested in ways that reward a guess over an admission of not knowing, in the same way that a multiple-choice exam rewards guessing. A model that always answers scores better than one that sometimes says "I am not sure", so models learn to answer.

Agents reduce the problem without removing it. The checking step catches many errors because a tool result is real: the comparison site either shows a tariff or it does not. The failures that slip through are the ones no tool checks, such as a quoted statistic or a summary of a document it only skimmed. Three habits protect you: ask for the source of any number you will act on, ask it to say when it is unsure, and read anything that goes out under your name.

What credits pay for

Most agent services charge in credits, and beginners are often surprised that two similar-sounding tasks cost different amounts. Credits pay for three things: the tokens the model reads and writes, the tool calls it makes, and in some products the minutes a browser or a computer is running. A one-line question costs almost nothing. A research job that reads thirty web pages, writes a report and checks its own figures costs more, because every page read is tokens in and every draft is tokens out.

On Jobbit the Free plan gives 100 credits a day with no card, enough for a few real tasks, and Go is £7 a month for 2,000 credits and one small app kept live on your own domain. Our guide to how much AI agents cost explains credits and tokens across ChatGPT, Claude, Manus, Muse and Jobbit, and the honest way to compare plans.

Where you fit in

Anthropic's guide says human checkpoints are part of the design, not a sign the agent has failed. In practice you have four jobs, and none of them is doing the task.

1. Write the goal properly. What you want, what good looks like, what it must not do. Our guide on how to write prompts for AI agents has 30 examples.
2. Approve the plan when the task matters.
3. Answer the questions it stops to ask, and take over the browser when a site needs a human.
4. Check the result before you act on it.

Big jobs add one more layer. An agent can split a large task between several smaller agents, each with its own desk, and then combine their work, which is how a 200-page research job or a whole website gets done, as our guide to AI subagents describes. And when a job needs hands rather than tokens, Jobbit's agent briefs a vetted professional from its human network and holds your payment in escrow.

What this means in practice

ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse all run some version of this loop, with the fuller tool sets on their paid plans. The loop is the same; what differs is how much of it you can see and steer.

• You can see the plan before it runs. Jobbit's plan mode shows the steps and waits for your yes.
• You can watch the browser. The agent browser shows every page it visits, and you can take the controls.
• You can steer mid-run. Send a message while it works and it adjusts, or fork the chat to try a second approach.
• You can give it a memory. Projects hold your instructions, files and secrets so every task starts informed.
• You can set it running on a schedule. Scheduled tasks repeat the loop without you.
• You can hand off to a person. When the loop reaches a job that needs a human, the same agent finds and briefs one.

A five-minute experiment that shows the loop

• Minute 1: open Jobbit and start a free account.
• Minute 2: ask it: "Compare the three cheapest ways to send a 5 kg parcel from Leeds to Glasgow this week, with links, and tell me which one you would pick and why." A goal, a definition of done and a request for sources.
• Minutes 3 and 4: watch the plan appear, then the browser open the carriers' sites. Notice the checks when a price needs a postcode.
• Minute 5: read the result and click one link to confirm a price. You have just seen all five steps, and you know what to check next time.

Frequently asked questions

How does an AI agent work step by step?

It reads your goal and decides what done looks like, writes a short plan, uses tools such as a browser and your files to carry out each step, checks each result against the plan, and reports back, stopping for your approval before anything involving money or messages.

What is the difference between an AI model and an AI agent?

The model is the engine that reads and writes language, such as Claude or GPT. The agent is the model plus tools, a plan, a checking loop and a way to stop and ask you. A model answers; an agent acts.

Does an AI agent learn from me?

Within a task it remembers everything you have said. Between tasks it only remembers what is saved in instructions, files or a memory feature. Anything you type twice belongs in saved instructions.

Why do AI agents make mistakes?

Because the model inside produces the most likely text, not guaranteed truth, and research shows models are trained in ways that reward guessing. Tools and checking catch many errors; you catch the rest by asking for sources and reading what goes out.

Can an AI agent browse the internet?

Yes, with a browser tool it can visit websites, fill in forms and read pages the way you do. Some websites block agents, and on Jobbit you can watch the browser and take over when a site needs a person.

What are credits in AI agents?

The unit most services charge in. They pay for the text the model reads and writes, the tools it calls and sometimes the time a browser runs. Jobbit's Free plan gives 100 credits a day with no card.

Can I watch an AI agent work?

On Jobbit you can: it shows the plan, the browser and each step, and you can stop it, steer it or take control at any point. 100 free credits a day, no card, and Go is £7 a month for one small app on your own domain.