How AI agents work: a goal, tools, memory and a loop of checks.

An AI agent is software built around a language model that is given a goal rather than a single instruction. It plans a step, uses a tool such as an inbox or a database, checks what happened, and repeats until the goal is met or a person needs to step in. That loop is the whole idea.

What is the difference between an AI agent and a chatbot?

A chatbot answers a message and waits for the next one. An agent takes a goal and works towards it in other software: it reads an email, looks up a record, drafts a reply and sends it. Anthropic draws the line between workflows, which follow predefined code paths, and agents, where the model directs its own steps and tool use.

Most useful business systems sit somewhere between the two. A fixed workflow with one AI step in the middle, such as classifying an enquiry, is often the right answer. Our guide to which business tasks to automate first covers that end of the scale.

What are the parts of an AI agent?

Every working agent has four parts: a goal, tools, memory and checks. Anthropic describes the basic building block as a language model enhanced with retrieval, tools and memory. The checks are what turn that building block into something a business can trust with real work.

• The goal: one job, described clearly, with a definition of done. Reply to every new lead within the hour is a goal. Help with sales is not.
• Tools: the actions the agent may take, each with a name, a description and the inputs it needs. Send an email, look up a customer, create a task. If a tool is not on the list, the agent cannot do it.
• Memory: what it knows between steps. In the short term, that is the conversation so far and the result of every tool call. In the longer term, it is records it can look up, such as which leads were already contacted. Looking things up in documents is called retrieval, which we explain in what is RAG.
• Checks: rules that sit outside the model. Was the email delivered? Does the draft promise something the business never offered? Does this action need a person to approve it first?

How does an AI agent decide what to do next?

It works in a loop of plan, act and check. The model looks at the goal and what it knows so far, picks the next step, calls a tool, reads the result and decides again. It stops when the goal is met, when a check fails, or when it reaches a limit you set.

Plan, act, check. Then round again.

The acting happens through tool use, also called function calling. Claude's documentation describes it plainly: the model decides when to call a tool based on the request and the tool's description, then returns a structured call that your application runs. The result goes back to the model, which uses it to choose the next step. So the model never touches your systems directly. Your code does, and your code can say no.

The check is the step people skip. Anthropic's guidance says agents need ground truth from the environment at each step, such as the actual result of a tool call, to judge their progress. It also recommends stopping conditions, such as a maximum number of iterations. In plain terms, the agent should look at what really happened rather than what it expected, and it should never be able to loop forever.

Some checks belong to people. An agent can pause at a checkpoint and wait for approval before it sends, pays or deletes anything. That pause costs a few seconds of someone's attention. Skipping it can cost a customer.

When should you not use an AI agent?

Skip the agent when the steps are fixed and the rules are clear. A plain workflow or a few lines of code will be cheaper and easier to test. Anthropic notes that agentic systems trade latency and cost for better task performance, and that their autonomy brings the risk of compounding errors.

• The input is structured and predictable, such as a form with fixed fields.
• The same input must always give the same output, such as an invoice total or a tax calculation.
• A mistake would be expensive and nobody is available to review the work.
• You cannot yet describe what a good result looks like, so nothing can check it.

Agents earn their place where the input is messy: free-text emails, varied documents, conversations that can go several ways. That is where fixed rules break and judgement helps.

What does a real AI agent look like?

We built an AI email-outreach system for ETRL that sends emails to leads automatically. It handles the writing of each email, checks that it was delivered, and replies when a lead answers so the conversation keeps moving. It has generated more than 500 leads and saved more than 20 hours of work.

Seen through the four parts, the design is not exotic. The goal is narrow: start conversations with leads and keep them going. The tools are the ones a person would use by hand: something to write and send with, a way to confirm delivery, a way to read replies. Memory is the record of who has been contacted and what was said. And checking delivery is one of its jobs, not an afterthought.

That is the shape to look for in your own business. A job with clear steps that someone repeats every day, where the input varies enough that a fixed script struggles.

How do you start with an AI agent?

Start with one job, not a platform. Write down the goal, the tools a person uses to do it today, what they need to remember between steps and how they know it went right. That single page is the first draft of an agent design, and it often shows that a simpler workflow would do.

Our guide to AI agents for a growing business is the next step. It covers choosing that first job, proving it on real data and the guardrails that make it safe to switch on.

If you have a job in mind, custom AI and agents is how we build them, and AI and automation shows where agents sit next to simpler workflows. Or tell us about the job and we will say plainly whether it needs an agent at all.