Agentic AI: what this buzzword really means

Agentic AI describes a way of making artificial intelligence work in several steps: it plans, uses tools, checks what it gets back and corrects course, instead of answering in a single shot. Behind the buzzword sits a simple and useful idea, along with plenty of marketing labels worth learning to decode.

Where the word “agentic” comes from

“Agentic” comes from agency: the ability to act on your own behalf. Psychologists used the word long before AI did, but in the last couple of years it has spread to product launches, investor decks and job ads across Silicon Valley and the City of London.

One milestone comes up again and again. On March 20, 2024, Andrew Ng, the researcher who founded DeepLearning.AI, published a letter in The Batch titled “How Agents Can Improve LLM Performance” (accessed October 22, 2026). He contrasted two ways of using a large language model: asking it to write something in one go, or letting it work iteratively, the way a person outlines, drafts, rereads and revises. He called the second approach an “agentic workflow”.

So the term first describes a way of working, not a product. That is what makes it both handy and vague.

A definition that holds up

Here is a working definition: an AI system is agentic when a language model decides, at least in part, what to do next to reach a goal, acting on its environment through tools and taking the result of each action into account.

Three ingredients stand out:

1. A goal, rather than a single question.
2. Tools: web search, reading files, sending email, running code. Technically, the model requests them through tool calling.
3. A loop: the system looks at what its actions produced before choosing the next step.

Anthropic, the company behind Claude, adds a useful nuance in “Building Effective Agents”, published December 19, 2024. It groups all these systems under the “agentic” umbrella, then splits them in two: workflows, where models and tools are orchestrated through code paths defined in advance, and agents, where the model directs its own steps and its choice of tools.

In other words, agentic AI is not an on/off switch. It is a dial, running from a pre-planned chain of steps to an agent that improvises its own plan.

Assistant, workflow, agent: a spectrum, not a box

The table below places the main forms on that dial. Our article on the differences between AI agents, chatbots and assistants goes further with more examples.

Form	Who decides the steps?	Tools	Example
Chatbot or classic assistant	You, message by message	Few or none	Rewording an email
Agentic workflow	The developer, in advance	Yes, in a fixed order	Sorting incoming invoices, then logging them in a spreadsheet
AI agent	The model, as it goes	Yes, picked on the fly	Comparing three hotel offers and building a table
Multi-agent system	Several models splitting the work	Yes, by specialty	One agent researches, another writes, a third reviews

The further down the table you go, the more flexible the system becomes, and the harder it is to predict and control. For a full introduction to agents themselves, read our guide what is an AI agent.

The four building blocks of agentic AI

In his March 2024 letter, Andrew Ng described four patterns that make up most agentic systems. They remain a good lens.

Reflection

The model reviews its own work and looks for ways to improve it. This is called self-reflection. It works much better when the system has a concrete signal, such as a failing code test, than when it grades itself blindly.

Tool use

The model can search the web, run code, check a calendar or query a database. That is what turns words into actions. Protocols such as MCP now standardize these connections, a topic we unpack in MCP explained simply.

Planning

Faced with a broad goal, the system breaks it into subtasks and handles them one by one. This is task decomposition. How an agent chains planning, action and checking is the subject of our article how do AI agents make decisions?.

Multi-agent collaboration

Several agents, each with a role, share the work and compare results. That is the principle of a multi-agent system.

Where you already meet agentic AI

You do not need a research lab to run into agentic AI. As of October 22, 2026, several mainstream assistants offer an agent mode, according to their makers’ websites: ChatGPT, Claude, Perplexity and Le Chat, for instance, can chain searches, open pages or produce documents from a single goal.

Assistant	Free	Paid from	Web search	Agent mode	Voice	Images
ChatGPT	Limited	$8/mo	Yes	Yes	Yes	Yes
Claude	Limited	$20/mo	Yes	Yes	—	—
Perplexity	Limited	$10/mo	Yes	Yes	—	Yes
Le Chat (Mistral Vibe)	Limited	$5.99/mo	Yes	Yes	Yes	Yes
Manus	Limited	$20/mo	Yes	Yes	—	—

For businesses, agentic AI often takes the shape of automations: a no-code tool connects an inbox, a language model and a spreadsheet, then runs the chain every time a new message arrives. Developers use dedicated frameworks instead. Our AI agent framework comparison sorts them by type, license and required skill level.

The buzzword problem: how to spot “agent-washing”

Because the term sells, it sometimes gets stuck on products that are nothing new: a decision-tree chatbot, a classic automation or a “smart” form. Three questions help you sort things out.

• Does the system choose its own steps? If it always follows the same path, it is an automation, possibly a useful one, but not an agent.
• Does it use real tools? An assistant that only generates text is not acting on anything.
• Does it check the results of its actions? No observation loop, no agentic behavior.

Check yourself A piece of software automatically sends the same AI-written follow-up email three days after every quote. Is that agentic AI?

Not really. The AI writes the text, but the steps and their order are fixed in advance, and the system decides nothing based on results. It is an automation with generative AI inside. It would become agentic if the system read customers’ replies and chose what to do next: follow up again, offer a call or hand over to a human.

What agentic AI changes, and where it falls short

The benefit is real. An AI that searches, compares, drafts and reviews its own work saves time on multi-step tasks, where a classic assistant made you steer every message yourself.

But every extra step is another chance to go wrong. An early mistake carries through the rest of the chain. An agent that reads web pages can also stumble on hidden instructions and be hijacked: this is called prompt injection. And more steps mean more calls to the model, so more time and more cost. Our article on autonomous AI agents looks at how far you can reasonably let them act alone.

Before handing a real task to an agentic system, take a moment to assess what could go wrong: