The short answer: nobody agrees, and the disagreement is not trivial. Three of the most frequently cited sources on this question define the terms in ways that cannot all be true at once. One treats the split as autonomy against predefined code paths, one treats it as one agent against several working together, and one treats "agent" as the whole category and does not use "agentic AI" as a separate thing at all. When a vendor tells you their product is agentic rather than an agent, the claim is not meaningful until you know which of these they mean, and most of the time they have not decided.
That sounds like pedantry. It is not, because the three definitions imply completely different products, different failure modes, and different things you are on the hook for.
The three definitions actually in use
The distinction is autonomy over control flow. Anthropic's engineering write-up draws the line between two kinds of system, and puts the boundary at who decides what happens next. It describes workflows and agents this way: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." Note what this version does not care about: how many agents there are, or how clever any of them is. A single model choosing its own next step is an agent. A dozen models on rails are a workflow.
The distinction is one agent against many. MIT Sloan's explainer reports a different line, and is candid that the usage is unsettled. In its account of the term, agentic AI is defined as systems that incorporate multiple, different agents orchestrating a task together, with a marketplace of agents representing both sides of a negotiation given as the example. The same piece notes that most people still use the two terms interchangeably. Under this definition Anthropic's single autonomous agent is not agentic at all, because there is only one of it.
There is no distinction worth drawing. Google Cloud's explainer does not split the terms. It defines the category outright: AI agents "are software systems that use AI to pursue goals and complete tasks on behalf of users. They show reasoning, planning, and memory and have a level of autonomy to make decisions, learn, and adapt." Agentic is left as an adjective describing that, not a separate class of thing.
So one source puts the boundary at control flow, one puts it at cardinality, and one declines to put it anywhere. These are not shades of the same idea. A product that is an agent under the first definition can fail the second, and the third would call both of them the same word.
Why the confusion is load-bearing for a buyer
The terms are doing sales work. "Agentic" arrived as a way of saying "more advanced than the last thing", and because no definition constrains it, it costs a vendor nothing to claim. A chatbot with two API calls bolted on can be described as agentic by someone using the Google Cloud framing loosely, and nobody has technically lied.
The questions that survive the terminology are the ones worth asking instead:
- Who decides the next step, the code or the model? This is Anthropic's line, and it is the one that predicts behaviour. A system on predefined paths does what you told it and fails predictably. A system choosing its own path handles cases you did not anticipate, and fails in ways you did not anticipate either.
- What can it do without asking? Autonomy over control flow is a different question from autonomy over consequences. A system can pick its own steps and still be required to stop before anything irreversible. Most of the risk lives in the second question and most of the marketing lives in the first.
- What happens when it is wrong? Not whether it will be. A ten-step process that is 95% reliable per step is not 95% reliable overall, and the interesting property of a system is what it does with the error rather than whether it has one.
- Who is accountable for the output? If the answer is "the model", that is not an answer. Something has to be reviewable, attributable and reversible, and that is a product decision rather than a definitional one.
Notice that none of those four questions require you to settle what agentic means. That is the point. The vocabulary is unsettled and probably will stay that way for a while, and the useful move is to stop trying to place a product in a category and start asking what it does when it is uncertain.
Where we land, and what we call it
We do not use "agentic" as a product claim, because we would not be able to say which definition we meant. Sanaf is one AI employee that holds a job description per job, works in Slack or Microsoft Teams and the apps a business already runs, and stops before anything it cannot undo.
By Anthropic's definition it is an agent: it decides its own next step rather than running a path somebody drew. By MIT Sloan's it is not agentic, because it is one employee rather than a marketplace of them. Both of those are true at the same time, which is a reasonable illustration of why the terms are not carrying much weight.
What we would rather be judged on is the third question above. It starts in shadow mode and drafts rather than sends until it is promoted, approvals are set per kind of action rather than per step somebody remembered to add, and everything it did is on a record you can read afterwards. Those are properties you can check. "Agentic" is not.
The questions to ask a vendor instead
If you are comparing products and the category words are not helping, these five get you further than any definition will:
- Does the model choose its own steps, or does it follow a path somebody configured? Ask for the failure case, not the happy path.
- Which actions can it take without a person, and how is that list configured? "Human in the loop" with no detail means nothing.
- Does it draft before it sends, and can that be turned on for a trial period?
- What is the record afterwards, and can somebody who was not there reconstruct what happened?
- When it is uncertain, does it stop and ask, or does it guess and continue?
A product that answers those clearly is a known quantity regardless of what it calls itself. A product that answers them with the word "agentic" has told you nothing, and now you know why.