What does agentic mean?

Agentic means having agency: acting on your own initiative rather than being directed step by step. Applied to software, an agentic system is one that pursues a goal across multiple steps and decides each next action itself, based on what the previous action produced. The word is not new. It comes from psychology, where it describes a person acting as the origin of their own behaviour, and it was borrowed into AI because it names the property that distinguishes a goal-seeking system from a tool that waits for instructions.

Is agentic a real word?

Yes, and it long predates AI. It appears in psychological literature through the twentieth century, notably in work on human agency and in Stanley Milgram's description of an "agentic state", the condition in which a person sees themselves as an instrument carrying out another's wishes rather than as an actor with their own responsibility. The current AI usage is a borrowing from that literature, not an invention.

What is the difference between agentic and automated?

Automation follows a path someone defined. Agentic systems choose the path. A script that runs the same five steps every time is automation even if a language model performs each step. A system that reads a result, concludes the plan was wrong and tries something else is agentic. The distinction matters operationally rather than semantically: automation fails predictably and can be tested exhaustively, and agentic systems cannot, which is why they need supervision rather than test coverage alone.

What makes an AI system agentic?

Three properties together. It has a goal rather than an instruction. It can take actions in the world, usually by calling tools. And it decides its own next action based on the outcome of the last one. Remove the third and you have a workflow. Remove the second and you have a planner that cannot do anything. Systems marketed as agentic frequently have the first two and not the third.

Is agentic AI just a buzzword?

The word is overused and the underlying distinction is real, which is an uncomfortable combination. It has been applied to products that are chatbots with a function call, and that dilution is worth resisting, because the difference between a system that follows a defined path and one that chooses its own genuinely changes how you deploy, supervise and insure it. The buzzword complaint is fair. The concept it is attached to is not empty.

What is the opposite of agentic?

In psychology, the contrast is usually communal or, in Milgram's framing, the agentic state is contrasted with an autonomous state where a person owns their own decisions, which inverts the everyday intuition about the word. In software, the practical opposite is deterministic: a system whose behaviour is fully specified in advance. That is not a lesser thing. For most problems it is the correct choice, and it is easier to debug.

Agentic: What the Word Means, and the Test for Whether Something Is

Agentic means it decides the next step

Agentic means having agency: acting on your own initiative rather than being directed. In software it describes a system that pursues a goal across many steps and chooses each next action itself, based on what the last one produced. That is the whole definition, and it yields a single test that most products described as agentic do not pass.

By Thomas Prommer · Published 11 August 2026 · 6 min read

The definition

Agentic (adjective): acting as an agent; having and exercising agency. Of a system: pursuing a goal through actions it selects itself rather than following a sequence someone specified.

The technology industry borrowed the word rather than coining it, which is why it can feel simultaneously novel and oddly formal. It has a long history in psychology, describing a person acting as the origin of their own behaviour. Stanley Milgram used agentic state in his 1974 account of the obedience experiments to name something almost opposite to the modern connotation: the condition of seeing yourself as an instrument executing another person's will, and therefore not responsible for the outcome. Anyone reaching for the word to mean "empowered" is using a sense the older literature would not recognise.

The one-question test

Who chooses the next step?

If the sequence is fixed before the system runs, it is automation, and putting a language model inside each step does not change that. If the system reads the result of one action and decides what to do next on that basis, including deciding that its plan was wrong, it is agentic.

This is not pedantry, because the two fail differently. Automation fails predictably and can be tested exhaustively: same input, same path, same bug. An agentic system's failure surface is the set of paths it might choose, which you cannot enumerate. That is precisely why one needs test coverage and the other needs supervision. Calling the first thing agentic to make it sound modern imports the second thing's operational burden into a system that did not need it, or worse, skips the burden entirely.

Three properties, and the one that gets dropped

A goal rather than an instruction. "Make the tests pass" instead of "edit line 42".
The ability to act. Tools, commands, API calls. A system that can only produce text is advising, not acting.
Self-directed iteration. It observes the result and chooses again.

Products marketed as agentic almost always have the first two. The third is the expensive one to build and the risky one to ship, and it is the one that makes the word accurate.

How the term is misused

Chatbot with function calling. Genuinely useful. Still turn-by-turn, still directed by you, not agentic.
Multi-step prompt chain. The steps were chosen by a developer. That is a workflow, and often the better design.
Anything with "AI" rebranded. The 2026 equivalent of appending "cloud" to a product name in 2013. AP News covered the term's rise as tech's newest buzzword for a reason.
Agentic used as a compliment about people. Common in management writing, and it collides with the older psychological sense in a way that would make Milgram uncomfortable.

Why the distinction is worth keeping

Because it predicts what you have to do to operate the thing. A deterministic system gets tests, a runbook and alerting on known failure conditions. An agentic system also needs a way to detect that it has silently stopped making progress, a recovery path that does not destroy work in flight, and a bounded blast radius, because you cannot enumerate what it might decide to do. Those are different budgets and different skills. A team that buys the label without the practice ends up with the risk profile of the second and the operational maturity of the first.