What Is Agentic AI? A Plain-Words Definition With Examples

Agentic AI is AI that is given a goal rather than a question, and works towards it by deciding what to do next, using tools to do it, looking at what happened, and carrying on until the goal is met or it is stuck. A chatbot that writes you an email draft is not agentic. A system that reads your inbox, decides which messages need a reply, drafts them, files the rest and tells you what it did is. The difference is not how clever the model is. It is whether the software takes actions in the world on your behalf, several in a row, without you pressing the button for each one.

“Agentic” is an adjective, and that matters. It describes a degree of independence a system has, not a separate kind of technology. What a single AI agent is, in one paragraph, is in the glossary; this page is about the wider idea and how the word is used.

Where the term comes from

The word is older than the current products. In a 2023 paper on the harms of increasingly agentic algorithmic systems↗ (opens in a new tab), Chan and colleagues argued that agency is not a yes-or-no property but a matter of degree, and named four things that make a system more agentic: underspecification (it is given a goal without being told how), directness of impact (its outputs act on the world without a person in between), goal-directedness, and long-term planning.

The word reached everyday use in 2024. Andrew Ng’s March 2024 letters in The Batch on agentic design patterns↗ (opens in a new tab) described an agentic workflow as one where a model is prompted many times, planning, drafting, reviewing and revising, instead of writing its final answer in one pass, and named four patterns: reflection, tool use, planning and multi-agent collaboration.

Later that year Anthropic’s guide to building effective agents↗ (opens in a new tab) admitted that “agent” can be defined in several ways, from fully autonomous systems that run for long periods to prescriptive ones that follow predefined steps, and used “agentic systems” as the umbrella for all of them. Within that umbrella it separates workflows, where code decides the steps, from agents, where the model directs its own process and tool use.

So there is no single official definition, and vendors stretch the word to cover almost any product with an AI feature. The useful reading is the one the sources share: the more a system decides its own next step and acts on it, the more agentic it is.

The five properties people mean

When someone calls a system agentic, they usually mean some mix of these five. A system does not need all of them to earn the word, but the more it has, the more the word fits.

• A goal, not a prompt. It is told what should be true at the end — “every failing test passes” — rather than what to write.
• Planning. It breaks the goal into steps and decides the order, and it can change the plan when something turns out differently.
• Tool use. It can search, read files, call APIs, run commands or fill in forms. Without tools a model can only produce text; with them it can change things.
• Acting. It carries out those steps itself, one after another, without a person approving each call.
• Adapting. It reads what each action returned, including errors, and uses that to decide what to do next, going round a loop until it is done or blocked.

That loop — plan, act, observe, check, repeat — is the engine of every agentic system. Each step, and how each one goes wrong, is walked through in the steps an AI agent takes to complete a task.

Levels of autonomy

Because agency is a matter of degree, “agentic” covers everything from an assistant that proposes each action for a person to approve, to one that acts on its own and is only reviewed afterwards. Most teams use several levels at once: the same assistant may fix lint errors unattended and wait for a person before it emails a customer. How to choose the level for each kind of job — by how easily it is undone, how far a mistake spreads and how easily it can be checked — is the subject of human in the loop vs fully agentic AI.

Eight examples across different jobs

1. Software. A coding assistant is asked to fix a failing test. It runs the suite, reads the error, finds the code, edits it and runs the tests again. Claude Code’s documentation describes this as an agentic loop↗ (opens in a new tab) of gathering context, taking action and verifying results, repeated until the task is done.
2. Research. Asked a broad question, a lead agent splits it into parts and sends several sub-agents to search at once, then combines what they found. Anthropic describes its own Research feature working this way in how it built a multi-agent research system↗ (opens in a new tab).
3. Customer support. An assistant reads a refund request, looks up the order, checks it against the refund policy, and either issues a small refund within a set limit or drafts a reply for a person to approve.
4. Operations. Overnight, an assistant groups the day’s error log, checks the task list for each problem, and files the new ones as bugs with the file and line they point to.
5. Finance. An assistant matches incoming invoices against purchase orders, flags the ones that do not match, and prepares the rest for a person to approve for payment.
6. Sales. An assistant researches each new lead from public sources, fills in the company record, and drafts a first email for the account owner to send.
7. Personal admin. Given “find a time for the four of us next week”, an assistant reads the calendars it has access to, proposes slots and sends the invitation once you say yes.
8. Browsing. An assistant working in a web browser fills in a long form from a document you gave it, stopping before the submit button for you to check.

Notice what they share: a goal, tools that reach real systems, and a point where a person steps in before anything expensive or irreversible happens. More jobs you can hand over, with the review call for each, are in AI agent task examples.

What agentic AI is not

• Not just a chatbot. A model that answers questions, however well, is generative AI. It becomes agentic when it can act. The comparison is set out in agentic AI vs generative AI.
• Not a new kind of model. Agentic systems are built on the same large language models as chat assistants; the agency comes from the tools, the loop and the permissions wrapped around the model.
• Not automation with a new name. A fixed script that runs the same steps every time is automation. It becomes agentic when the model chooses some of the steps.
• Not the same as autonomous. An agentic system can still wait for a person at every consequential step. Autonomy is a setting you choose per job, not a property of the technology.
• Not unsupervised by default. The safe versions act inside permissions someone set, and leave a record someone can read.

What it changes for the people around it

Once software acts rather than answers, three questions that never came up with a chatbot become everyday ones. What is it allowed to do? What did it actually do? And where does the work it was given come from, and where does it report back? A transcript answers none of them well: it is long, mostly tool output, and gone when the session ends.

A task board answers all three. On fenbs an AI assistant is a member of the board: it acts as the person who connected it, holds that person’s role narrowed by the scopes ticked when it was approved (read, write, comment), and every change it makes is recorded in History as, for example, “Claude via Sam”. It takes its work from the tasks on the board and reports on them — a comment, a move to Completed, a Testing status saying what was checked. A person can pre-approve a task with “Let AI do this”, and an assistant picks it up with fenbs_next_approved_task; the card then says “AI done · check it” until a person confirms it. Revoking its token stops it at once without touching anyone else’s sign-in.