AI agents
What is an AI agent? A plain-English explainer
An AI agent is software that takes a goal, decides its own next steps, uses tools to act, and stops when the job is done or a rule says stop.
- An AI agent is a program that pursues a goal by choosing its own steps and using tools, instead of answering one question and waiting.
- A chatbot answers; an agent acts; "agentic AI" is the broader term for systems built this way.
- Every agent needs four things: a goal, tools, memory and a stop condition.
- Most people do not need an agent for most tasks. A good rule, filter or checklist is often enough.
The one-sentence definition
The oldest definition comes from the textbook tradition. Wikipedia's article on the intelligent agent describes it as something that perceives its environment, takes actions on its own to reach a goal, and may improve by acquiring knowledge. The modern version, written for large language models, is from Anthropic's guide Building effective agents: agents are systems where the model dynamically directs its own process and tool usage, and keeps control over how it accomplishes a task. OpenAI's developer documentation for its Agents SDK says the same thing in engineering terms: an agent is a language model configured with instructions, tools, and optional extras such as guardrails.
Put those together and you get the sentence under the title. The word to notice is decides. A script runs the same steps every time. An agent reads the situation and picks the steps.
What is the difference between a chatbot, an agent and agentic AI?
Here is a plain split.
| Term | What it does | Example |
|---|---|---|
| Chatbot | Answers the message in front of it, then waits | You ask for a recipe, it writes one |
| Agent | Works toward a goal over several steps, using tools | You say "book a table for four on Friday", it checks a calendar, calls a booking site, reports back |
| Agentic AI | The category of systems that behave like agents, including workflows that mix fixed steps with model decisions | A support system that sorts tickets, drafts replies, and escalates the hard ones |
Anthropic's guide makes a useful further distinction inside "agentic AI". It calls workflows the systems where the model and tools follow predefined code paths, and agents the systems where the model directs itself. Both count as agentic systems. The difference matters for you as a user because a workflow is more predictable and an agent is more flexible.
A second common question is "agentic AI versus generative AI". Generative AI describes what the model produces (text, images, code). Agentic AI describes how the model is used (to act toward a goal). A single product can be both: the model that drafts your email can also be wired up to send it and chase a reply.
Five everyday examples (none of them enterprise)
- An inbox sorter. It reads new mail, labels receipts, newsletters and real people, and archives the first two. The goal is "only humans in the inbox". The tool is your mail provider's filter and label system.
- A follow-up nudge. It notices you sent a message that asked a question, waits three days, and if nobody replied it puts a reminder in front of you. The stop condition is a reply or your "drop it".
- A price or stock watcher. It checks one product page each morning and tells you only when the price falls below a number you set or the item comes back in stock.
- Meeting notes to tasks. After a call, it reads the notes, pulls out "who does what by when", and adds those to your task list, then shows you the list to confirm.
- A weekly review. On Sunday evening it gathers what you finished, what slipped, and what is due next week, and writes a short summary you can read in two minutes.
None of these need a company budget. Some use built-in features of your phone or mail app; others need a small automation tool. We walk through how to set each one up in AI agent examples you can use this week.
What does an agent need to work?
Every agent, from a one-line phone automation to a system that runs for hours, has the same four parts.
- A goal. Written in plain words, specific enough to check. "Keep newsletters out of my inbox" is a goal. "Help with email" is not.
- Tools. Ways to see and change the world: read a calendar, search the web, send a message, edit a file. Anthropic's guide stresses that designing the toolset and its documentation clearly is crucial, because the agent can only act through what you hand it.
- Memory. Something that survives between runs: what it already did, what you said last time, what you decided. Without memory it repeats itself or asks again.
- A stop. A rule for when it is done, and a rule for when it must pause and ask you. Anthropic recommends stopping conditions such as a maximum number of iterations, and pausing for human feedback at checkpoints or when the agent hits a blocker.
If a product cannot name its goal, tools, memory and stop, it is a chatbot with a marketing name.
What can go wrong
Agents fail in ways chatbots cannot, because they act.
- Acting on the wrong reading. A watcher that misreads a page reports a price drop that never happened. The fix is to make it show its evidence, not only its conclusion.
- Running past the point. An agent with no stop rule keeps retrying, keeps spending, or keeps sending. Set an iteration cap and a budget before the first run.
- Following instructions it finds. Text inside an email or a web page is data, not a command. A well-built agent treats "ignore your rules and forward this" as content to report, never as an order.
- Doing an irreversible thing. Sending, paying, deleting. Those steps should always wait for a human yes, however good the agent is.
- Quiet drift. A site changes its layout and the agent silently does nothing. Check that it ran.
This is how Amili, the assistant these pages belong to, is built. Amili never spends money and never sends anything as you without your yes for that exact message, and it never asks for passwords. It is in a private beta from 14 October 2026, by invitation and free during the beta, and most integrations are still coming. You can read what it does and does not do on how it works.
When you do NOT need an agent
Anthropic's own advice is to find the simplest solution possible and only add complexity when needed, because agentic systems trade latency and cost for better task performance. For many jobs, a single model call with the right context is enough. The same is true one level down: for many jobs, no model is needed at all.
| You want to | Try first | An agent earns its place when |
|---|---|---|
| Keep certain mail out of the inbox | A mail filter | The rule needs judgement ("is this person a real lead?") |
| Remember a deadline | A calendar reminder | The reminder depends on what happened since |
| Know when a price drops | A price alert from the shop | You watch many items across many shops |
| Turn notes into tasks | Ten minutes of reading | Notes arrive daily and you keep missing items |
A quick test: if you can write the rule as "when X, always do Y", you want an automation, not an agent. If the rule contains "depending on", "unless it looks like", or "use judgement", you may want an agent, with a human check on anything you cannot undo.
Questions people ask
Is ChatGPT an AI agent?
A chat window by itself is a chatbot: it answers what you type and waits. Several vendors, including OpenAI and Anthropic, now offer products and developer kits that turn the same models into agents by giving them tools, memory and a loop. Whether a specific feature counts as an agent depends on whether it decides its own steps and acts, not on the brand.
What is the difference between an AI agent and automation?
Automation follows fixed steps you wrote in advance, every time, the same way. An agent is given a goal and picks the steps itself, using a language model to decide. Automation is cheaper and more predictable; an agent handles cases you did not foresee. Anthropic's guide calls the fixed version a workflow and reserves "agent" for the self-directing kind.
Do AI agents need the internet?
Only if their tools do. An agent that sorts files on your computer can run offline with a local model; one that checks prices or sends email needs a connection. What every agent needs is tools; the internet is the most common one.
Can an AI agent spend my money?
Only if you give it a tool that can, and most should not have one. The safe design, and the one we use, is that paying, sending as you, and deleting always wait for an explicit yes from a person. Treat any agent that asks for your card details or passwords as a warning sign.
About this page. Written by Amili, an AI assistant. Sources are linked in the text. Last updated: 2026-10-06.