Generative AI vs AI Agents vs Agentic AI vs Skills, Explained for Beginners

Four terms, one machine

Generative AI, AI agents, agentic AI, skills. People use them like they mean the same thing. They don’t, and most explainers teach them as a ladder. It isn’t a ladder. It’s one machine.

Here’s the short version.

1. Generative AI is the engine. It makes text, images and code.
2. An agent is that engine in a loop. It uses tools and keeps going until the job’s done.
3. A skill is a manual the agent pulls off the shelf for one job.
4. Agentic isn’t a thing at all. It’s a dial: how much the AI decides on its own.

I’m a data consultant and auditor. To keep this concrete, I’ll use one made-up example: a Diwali gift list, twelve relatives, a budget of 15,000 rupees. None of that is real data. It’s just a job to hang each term on.

Generative AI is one box: your prompt goes in, new content comes out, then it stops. For the gift list it gives you wishes and gift ideas, but no stock check, no prices, no delivery. It’s a brilliant friend on the phone: great advice, but its hands aren’t on your keyboard.

Give that friend your laptop and you have an agent. OpenAI’s own guide defines agents as systems that independently accomplish tasks on your behalf, and then lists what doesn’t count: simple chatbots and single-turn models. The picture changes from a line to a loop: the model picks the next step, uses a tool like a search or a spreadsheet, checks the result, and goes round again until the job’s done or it hands back to you. If a gift is out of stock, it swaps it. You never gave it that step. It decided.

Where the label hides the decision

That’s the line that matters: who decides the next step? If a person wrote the step in advance, that’s a workflow, not an agent. It looks smart, but someone else did the deciding.

And that’s where a lot of marketing hides. The word “agent” now gets stuck on everything. Picture an FAQ chatbot relaunched as an “AI agent”: same chatbot, new label. Gartner, the research firm, has a term for this: agent washing, which means rebranding an old chatbot or automation as an agent without adding real agent capabilities. Gartner estimates that of the thousands of companies selling agentic AI, only about 130 are the real thing.

A simple way to hold it: a chatbot talks, an agent acts, and a workflow follows a script someone else wrote.

Then there’s “agentic”, which carries two meanings. Imagine handing the gift list to your cousins instead of an AI. One way, you split the work: one finds the gifts, one tracks the budget, one sorts out delivery, and an elder sister keeps all three in step. That’s the first meaning: a team of agents with a manager.

The second way is just one cousin, and the question is how much you let them decide. You can say “send me every option before you buy anything”, or you can say “here’s 15,000, sort it out, tell me when it’s done”. Same cousin, same list. The second is far more agentic. Not how many helpers, but how much freedom.

Both meanings are in use. For you, freedom is the one that matters, because that’s where things go wrong. Agentic isn’t a fourth box. It’s an adjective that describes how a system behaves, not something you install. Agentic isn’t something you buy. It’s how much you let go.

The questions to ask a pitch

My auditor’s take: when a pitch says agentic, don’t argue about the definition. Ask two questions.

1. How much does it decide on its own?
2. Who checks its work before it spends money or sends a message?

If nobody can answer the second one, the dial is set too high.

For any tool that calls itself an agent, add one more: who decides the next step, the model or a script someone wrote?

The last two pieces are about what the AI knows and what it can reach. A skill is a folder that teaches an agent one job, opened only when that job comes up. Think of a new joiner who gets a list of the company’s manuals and pulls the payroll one only when a payroll question comes up. The agent sees only the labels, a name and a one-line description. When your request matches a label, it pulls the folder down and reads the instructions. Some folders also hold scripts it can run.

Mine screens 50 Indian stocks every morning: my rules in plain English, plus one Python script. I type “run my swing screen” and it does the rest. Skills also aren’t tied to one app anymore. They became an open standard in December 2025, and ChatGPT and Codex now read the same SKILL.md format the same way. It’s a format, not a feature.

But a skill only tells the agent how. It doesn’t give it a way into your stuff. That’s MCP, the Model Context Protocol. Anthropic released it in late 2024, and ChatGPT and others support it now. The official site calls it a USB-C port for AI applications. Claude’s Help Center puts it in one line: “MCP connections give Claude access to tools, while skills teach Claude how to use those tools effectively.” MCP gives it the keys. The skill tells it what to do once it’s inside.

The test to run this week

Five pieces, and most people only need two. Answer these four questions in order and stop at the first one that does the job.

1. Is the job done when you get words or an image back? That’s a normal chat app. Most everyday jobs stop here.
2. Are you pasting the same instructions again and again? That’s a skill.
3. Does the AI need your information from another app? That’s a connector, and many of those are MCP under the hood. Check what it can read and what it can send.
4. Does the job take several steps, where the next one depends on what just happened? That’s an agent. Keep the approval prompts on, especially before anything that pays or sends.

For the gift list, the ideas and messages are question one, and the family rules are question two. That’s a chat app and one skill. The agent with a cart is optional, and only if you approve every payment. As Gartner put it: “Many use cases positioned as agentic today don’t require agentic implementations.” Even the people who build agents say start with the simplest solution and add complexity only when you need it.

So here’s the practical step. The third time you paste the same instruction into a chat, make it a skill. On a paid Claude plan, custom skills are now uploaded under Customize, then Skills. The menu moved after I recorded the video. Use the smallest piece that does the job, and add the next one only when it isn’t enough.