Skip to main content
HomeAboutSolutionsIndustriesContact

Generative AI, Agentic AI and AI Agents

Three terms separated by one variable: how much the system decides for itself. Getting them confused is why "agent" projects so often ship as chatbots.

Verdeshell Team · 6 min read · Last reviewed

Generative AI produces content when asked. An AI agent chooses actions and takes them. Agentic AI is the adjective for systems built that way — the property, not a product category.

The autonomy spectrum from generative AI to AI agentsA left-to-right axis of increasing autonomy. At the left, generative AI, where a person chooses every step. In the middle, agentic systems, which choose some steps for themselves. At the right, AI agents, which choose their own actions and take them. The variable is how much authority is delegated, not how capable the underlying model is.Generative AIYou choose every step“Draft this email”AgenticIt chooses some steps“Look it up, then draft”AI agentIt chooses and acts“Resolve this ticket”you decideit decidesMore autonomy delegated →Same model throughout. What changes is how much authority you hand over.
The axis is autonomy, not capability. Moving right means the system chooses more of its own steps.

Generative AI: you decide, it produces

You supply the instruction; the model produces content. Every step of the process is chosen by a person. Ask for a summary, get a summary. Ask again with different wording, get a different summary.

The system has no notion of a goal that persists past the current request, no ability to act on the world, and no way of knowing whether what it produced was any good.

This describes the overwhelming majority of what is deployed in businesses right now, including most things labelled as agents.

AI agents: it decides, then acts

An agent is given a goal rather than an instruction, and it selects its own actions toward that goal. The defining additions are tool use and a decision loop: it can call an API, query a database, read a file, run a command — and it decides which to do next based on what happened last.

That is the real line. A system that produces text describing what should happen is generative. A system that makes the thing happen, checks the result, and adjusts is an agent.

Everything else people list as agent features — memory, planning, multi-step execution — follows from having a goal and the ability to act on it.

Agentic AI: the property, not the product

Agentic is best read as an adjective describing how much a system decides for itself. It is a spectrum, not a category you either qualify for or do not.

A support tool that drafts a reply for a human to send is barely agentic. One that also looks up the order, checks the refund policy and issues the refund within a limit is substantially agentic. Same underlying model; different amount of delegated authority.

Framing it as a spectrum is more useful commercially too, because it turns the question from "is this agentic" into "how much authority are we comfortable delegating, and what happens when it is wrong" — which is the question that actually needs answering.

What is not an agent

A chatbot with a system prompt. Constraining the model's persona does not give it a goal or the ability to act.

A retrieval pipeline. Fetching documents before generating is a fixed sequence someone wrote; nothing is being decided.

Scripted automation with a model in one step. If the branches are written in advance, the model is filling in a slot, not choosing a path.

None of these are worse than an agent. They are frequently the better engineering choice — more predictable, cheaper, easier to test. They just are not agents, and calling them agents makes it harder to reason about what can go wrong.

The question to ask

What can this system do that nobody explicitly told it to do in this run? If the answer is nothing, it is not an agent — regardless of what the licence calls it.

And the follow-up: when it does something wrong, what stops it? Autonomy without a bound is the risk the word "agentic" tends to obscure.

More in Concepts

AI, Machine Learning and Generative AI: What the Words Mean6 min readRAG, MCP and Agents Solve Three Different Problems7 min readAI Terms, Defined8 min readA Learning Path for Building AI Agents6 min read

Want this built properly?

We design and run these systems for clients. Tell us the problem and we will tell you whether an agent is the right shape for it.