What Are AI Agents? Agentic AI Explained in Plain English

You've brought yourself up to speed on some of the AI tools. You can write a good prompt, you've probably settled on Claude or ChatGPT, and perhaps you've got an MCP or two installed (MCPs are small add-ons that connect Claude to other tools and services) and you're creating files and documents with them.

But nothing runs unless you start it. There's no automation and no follow-up, and if you want to repeat a process you have to start again with a freshly written prompt. Meanwhile the words "AI agent" and "agentic AI" are everywhere, and it's hard to find a straight answer about what either of them means.

Today's piece, I hope, should get you to the next stage of your AI journey. So, what is an AI agent, and what is agentic AI? Put simply, an agent is the model you already use, given a goal and some tools and left to work out the steps for itself, and agentic AI is the name for software that works that way. By the end you'll have a first agent to build, and it can be running this afternoon with a saved prompt and one MCP. If you're still on ChatGPT or haven't settled on a tool yet, most of this runs on Claude, and new accounts get a free week of Claude Code.

What is an AI agent?

An AI agent is a large language model (an LLM, the kind of model behind Claude and ChatGPT) running in a loop, with a goal and tools it can call. It decides the next step, uses a tool and reads what came back, and it repeats that until the goal is met or it gets stuck and needs you. In a chat, the loop is you, then the model, then you again. In an agent it's the model, then a tool, then the model, and only now and then you. You've stepped out.

A Teacher's Tech video on Claude Code splits working with AI into three levels. In chat you ask and the model answers, at the building level you direct each step, and at the agentic level you describe a goal and the model works out the steps. The video says going from chat to agentic is like the difference between micromanaging a brand-new employee and trusting an experienced one.

I think that's about right - my definition is broadly similar although I take great pains to emphasise that ultimately an agent is defined by the prompt - and many of my workflows are hardcoded prompts that direct procedures (upload and download from WordPress with an edit procedure in between being a very good example of a basic content agent).

What is agentic AI?

Agentic AI is the broader idea. It's software that takes a goal and pursues it through actions, instead of waiting for you to prompt it for the next step. The difference is in what happens after the first instruction: an agentic system reads the situation, works out the steps, uses tools to carry them out and checks the result, with very little help from you once it's been pointed at the goal.

That's the difference between answering you and acting for you. Ask a chat model about five competitors and you get a summary back, and there the job stops. Give an agentic system the goal of a competitor briefing and it fetches the pages, pulls the data it needs and drafts the briefing itself.

Anthropic uses "agentic systems" as its umbrella term, and it counts both workflows and agents under it. The terms get muddled mostly because vendors use them loosely. "Agent" sells better than "script", so plenty of things get called agents that aren't.

Agentic AI vs generative AI

Generative AI is reactive. You prompt it, it generates something (copy, an image, a draft email, a snippet of code) and the work ends at the response.

Agentic AI takes a goal and keeps going. It calls on tools, reads data and decides what to do next, until either the job is finished or it needs you to step back in. IBM's Martin Keen draws the same contrast in his video on the two: generative AI waits for a prompt, while agentic AI pursues a goal.

Underneath, both run on the same large language model, and it's the setup around the model that changes. In generative AI the model works on its own and the output is whatever lands in your chat window. In agentic AI the model is wired into a loop with tools and the goal you set, which means the output is action in real systems. The diagram below puts the two side by side.

How AI agents work: the agent loop

Every AI agent, whatever it's called, runs the same loop underneath. Google Cloud describes a five-step version of it, but for a beginner four words are enough: goal, think, act, observe. The diagram below shows the loop.

Goal

The goal is the thing you want done, set in plain words, and it can be quite specific. For example: "Find the five most-cited articles on agentic AI from October to December and pull the key claims into a side-by-side table." Everything the agent does after this gets checked against that sentence.

Think

The model breaks the goal into steps and writes out its working before it acts. This is called chain-of-thought reasoning. In Claude Code you'll see it as the text the model produces before it does anything.

Act

The model picks a tool and uses it. That might be a web search, a file read or a database query, depending on what you've given it.

Observe

The tool returns something. The model reads what came back, checks it against the goal and decides whether to move on to the next step or change course. An empty search result, for example, can send it back to try a different query. Then it loops. The loop runs until the goal is met, until the model decides it needs to ask you something, or until it hits a limit you've set, such as "don't run more than fifteen tool calls".

It's mechanically very simple. The clever bit is that the model is the one deciding what to do next, not you.

Google Cloud's longer version of the loop has five steps:

1. Perception
2. Reasoning
3. Planning
4. Action
5. Reflection

In the reflection step, a well-built agent notices when something has failed and adjusts its plan, whereas a badly built one just stops.

Of course there are limits to how well that works in practice, which I'll come back to - but the principles of 'agentic AI' are ludicrously simple at a foundational level.

AI agents vs agentic AI

Most people use the two terms as if they mean the same thing, and the broad definition above allows for that. At Houtini we split them more tightly, and this is our own framing rather than an industry definition.

An AI agent is one focused thing doing one job. It might find a flight at the right price, summarise yesterday's customer messages or run an SEO audit on a competitor. Each one is useful, but narrow.

Agentic AI is what you have when several of those agents work together under a coordinator that knows the wider goal. The coordinator hands the right work to the right agent, gathers the results and decides what happens next.

Where MCP fits: the tools an agent uses

An agent without tools is a chatbot, and the standard for connecting one to its tools is MCP (Model Context Protocol). Anthropic donated it to the Agentic AI Foundation (part of the Linux Foundation) in December 2025. MCP's own site describes it as "a USB-C port for AI applications". Each MCP server fronts one or more tools, and one agent can call many servers, as the diagram below shows.

On its own, the Gemini MCP we publish is a tool. Point Claude at it with a goal like "watch the news for new posts on agentic AI from Anthropic, Simon Willison, and Hamel Husain - bring me a weekly digest, with the original URLs" and you've got an agent. The Gemini MCP does the searching and reading, and Claude decides what to look for and writes the digest. Save that goal as a prompt and you can run it every week. This is also, I suspect, part of why some of the big software-as-a-service companies - the ones whose value sat in a unique UI and a nice website - have seen their valuations come under real pressure over the last few months.

A beginner needs three tools. The first is file access. Claude Code has it built in, and in the Claude desktop app an ordinary chat can now read and write files in any folder you attach. That came with the merge of Cowork (Anthropic's agent mode) into chat on 16 September 2026, which is rolling out to Pro and Max first. The second is web search. Claude has this built in too but I also lean heavily on Brave or Firecrawl MCP. The third is one tool for the job, such as the Gemini MCP, the seo-audit MCP, the DataForSEO MCP or Brevo MCP for email template management and campaigns.

Three tools is plenty for a first agent, and at this point you're mostly only constrained by your own imagination and knowledge of the available tools.

Workflows vs agents

Anthropic's December 2024 piece on building effective agents draws the line in two sentences:

So in a workflow you set the route in advance, and in an agent the model picks it as it goes.

For most of what you'll build in your first few weeks, workflows are the right answer. They're cheap to set up, failure conditions are more predictable and they're easy to fix and iterate on. You can read what they did when something goes wrong, use the LLM to give feedback so that it edits a prompt in your workflow, and iterate. It doesn't take long to start having wild ideas about what could be possible once you're past this point.

The more autonomous agent earns its keep when the task is open-ended: coding where nobody knows how many files need touching, research where the next step depends on what the last one turned up, or debugging where you can't predict the path. A daily content update probably doesn't need one.

Examples of AI agents at work

Let's look at a content agent I use to help me look after my various web projects. One of my saved prompts is called create-explainer-guide.md. It's a long markdown file with research steps, voice rules, image planning, an edit pass, AI detection gates and an upload pipeline. When I run it on a topic, Claude works through the steps in order. It researches, writes a draft, edits the draft against a voice rubric, captures images and preps the upload to the site. It pauses when it needs me to confirm something, then carries on. In plain English that's a content-production agent, and the only "code" involved is the markdown file. Naturally I write the narrative guidance, the introduction, the 'who is this for' and, after many months of refinement, my agents for content management work reliably, on a schedule and let me concentrate on the client work.

Most of the rest of our working week runs the same way. For research, we give the agent a goal like "build a synthesis on agentic AI from primary sources", and it fetches, reads, extracts and cross-checks against our own data before drafting anything, which a person then reviews and edits. Shopify work goes through the Shopify CLI (its command-line tool) inside Claude Code, so the agent handles theme updates, Liquid extensions, plugins and database queries by calling the CLI and editing the right files.

Charts and diagrams come out of the Gemini MCP's image tools in the same conversation, although Claude now makes interactive charts natively as well. Analytics means pulling and joining data across Shopify, GA4 and Search Console, where the agent does the extraction and the join and we do the interpretation. And with a skills library and a design-to-code path, the same loop turns a Figma frame or a written brief into a working page.

Your saved prompt is already an agent

The long prompt you keep pasting into Claude Code is already the spec for an agent - the one with the steps, the rules, the "don't do this, do that" instructions and the bit at the end where you tell it to ask clarifying questions. Save it and you've got one.

These days you save it as a skill. A skill is a folder with a SKILL.md file inside it. The file opens with frontmatter (a description is recommended; the name is optional and defaults to the folder name), then the steps, and the folder can hold extra files the agent reaches for, as the diagram below shows. Claude loads only the description at start-up and reads the body when the skill fits the task, which means long instructions cost almost nothing until they're needed.

Slash commands (the /name shortcuts you type in Claude Code) have been merged into skills. A skill and an old-style file in .claude/commands/ both give you /name, and the old file still works.

Anthropic launched Agent Skills in October 2025 and published them as an open standard in December 2025. ChatGPT and Codex, Gemini CLI, GitHub Copilot, VS Code and Cursor all support the format now. The Greg Isenberg and Remy Gaskell episode on building agents calls them SOPs (standard operating procedures, like a real company has). As an aside, our Voice Analyser MCP analyses your writing style via an XML sitemap and generates a SKILL.md for your agent to write in your tone of voice.

You don't need a new tool to make this jump, because the upgrade is writing down what you already do, with explicit rules, in a markdown file.

Making it run on its own

Once you've got a skill that works, the next rung is having it run without you typing the command. Claude can schedule it in the cloud or on your own machine, and Dispatch lets you start it from your phone when you're away from the desk.

A routine in Claude Code runs on Anthropic's servers (it's still a research preview, on the Pro, Max, Team and Enterprise plans). You configure it once with a prompt, one or more repositories and a set of connectors (Claude's built-in links to services like Slack or Google Drive), and then it runs on a schedule, from an API call or when something happens on GitHub. Because it runs in the cloud it keeps going when your laptop is closed, but it has no access to your local files. Its output lands in a branch, a pull request or a connector instead. A routine also acts without asking your permission first, and a green status in the run list only means the session started and finished without an infrastructure error. It doesn't mean the task succeeded, so glance at the output before you treat it as gospel.

A Desktop scheduled task runs on your own machine, which has to be awake with the Claude app open. It can read and write your local files, which makes it the right home for something like the weekly news digest, written to a folder every Monday morning.

Dispatch lets you send a task from the Claude phone app to your computer, which then does the work using your local files and tools. It's in limited beta on the Pro and Max plans and no longer available to new users, and like a Desktop task it only works while your computer is awake.

After these, the next step is code. The Agent SDK gives you the same loop and tools that power Claude Code, programmable in Python and TypeScript, and I'll leave that for another article.

When agentic AI is the wrong answer

If the work is fully predictable, repeats the same way every time, and the cost of a wrong answer is high, classic automation is cheaper, faster and more reliable than an agent. Anthropic's own advice points the same way: "add multi-step agentic systems only when simpler solutions fall short."

Bash scripts, Power Automate, Zapier and n8n all do that job well. The signal that you need an agent is when you find yourself writing endless if statements to handle exceptions and edge cases. That's the point where a model that can reason starts to pay for itself. Agentic AI also needs clean inputs. If the data is bad, the integrations are flaky and the goal is fuzzy, an agent won't save you, and it'll just produce a more eloquent failure than a script would have.

Agents fail more often than the demos suggest. On SWE-Bench Pro, a coding benchmark with a public leaderboard run by Scale AI, the top score in September 2026 was 61.5%, which is about six in ten. That's a lot better than nothing, but it's a long way from letting an agent work unchecked. Build review steps into your workflows, and don't run an autonomous agent against a production system without explicit guardrails. I might build in an output evaluation and 'sense checks' with Brave Search, or a second opinion from Gemini, depending on the project.

How to build your first AI agent

If I had to pick one first agent for a Claude Code reader, it would be a saved research-and-summary skill wired to the Gemini MCP. It takes four steps.

1. Install the Gemini MCP, using the setup guide if you haven't got it yet.
2. Write a saved prompt that asks what you want researched, asks three clarifying questions before starting, searches the web with the Gemini MCP, reads the top results, writes a structured summary to a research/ folder with every source cited, and finishes with three "what I'd do next" recommendations.
3. Save it as .claude/skills/research-topic/SKILL.md, with a one-line description in the frontmatter (the short block of settings at the top of the file).
4. Run /research-topic followed by your topic. Watch the loop run and see what happens!

Under Anthropic's definition this is a workflow, because the steps are set in advance. In the everyday sense of the word it's agentic, though. It follows a process, makes decisions and asks before it assumes, which is the Teacher's Tech video's test for agentic. And next week you can run it again on a different topic.

If you do nothing else after reading this article, do that. Enjoy!