Build Your First AI Agent, Part 1: What an Agent Actually Is (No Code, No Hype)

"Agent" is the most overworked word in business software right now. It is on pricing pages, in keynote slides, and in the subject line of at least one email sitting in your inbox this week. Ask five vendors what it means and you will get five answers, most of them shaped like whatever that vendor happens to sell.

This series is for owners and team leads who want to build one small, useful agent without writing code and without buying into a pitch. Before building anything, you need a definition sturdy enough to make decisions with. That is the whole job of Part 1.

A Definition You Can Actually Use

Two plain definitions from people who build these systems are worth keeping.

The first comes from Anthropic, the company behind the Claude models. In Building Effective Agents, published December 19, 2024, its engineers separate two things that usually get lumped together. A workflow is a system where the model and its tools follow a path someone defined in advance. An agent is a system where the model directs its own process and its own tool use, and keeps control over how the task gets done. (The model here is a large language model, or LLM, the technology behind ChatGPT, Claude, and Gemini.)

The second is shorter. Developer and writer Simon Willison, in a September 18, 2025 post arguing that the term finally had a definition agreed on widely enough to be useful, put it in one sentence: "An LLM agent runs tools in a loop to achieve a goal."

Put the two together and you get a working definition with four parts.

The Four Parts

A goal. Not a single question, but an outcome: get this quote request answered, reconcile these two lists, find every page on our site that still shows the old phone number.

Tools. Things the software is permitted to do beyond producing text. Read an inbox. Look up a row in a spreadsheet. Check a calendar. Search a folder. Save a draft. Without tools, a model can only talk. With them, it can act.

A loop. The model takes a step, looks at what came back, and decides what to do next. Then it does that again. This is the part that separates an agent from what came before it, because nobody scripted the order of the steps ahead of time.

A stopping point. Either the goal is met, or the agent hits a limit you set: a maximum number of steps, an action it is not allowed to take, or a checkpoint where a person has to approve before anything continues. Anthropic's guide describes both human checkpoints and stopping conditions, such as a cap on the number of iterations, as ordinary ways to keep control.

A goal, some tools, a loop, and a place to stop. Everything else you will hear about agents is detail layered on those four.

Chatbot, Automation, or Agent?

Most of what is sold under the agent label today is one of the first two. One question sorts them: who decides the next step?

A chatbot answers. You ask, it responds, and you decide what happens next. Useful, and we have written about how to get real work out of ChatGPT and similar tools, but not an agent.

An automation follows a script. When a form is submitted, add a row, send an email, create a task. Some automations now include an AI step, such as summarizing the form before it gets filed. The path is still fixed by whoever built it. This is what Anthropic calls a workflow, and the same guide points out that for well-defined tasks, that predictability is an advantage.

An agent is handed the goal and the tools and works out the path itself. If the first lookup comes back empty, it tries another. If the request is missing information, it notices and asks.

None of the three is better in general. They fit different jobs. If you already run marketing automation workflows, you own the middle category, and it may be all a given task needs.

What the Loop Looks Like

Here is a hypothetical example, invented for illustration and not drawn from any real company.

A service business gets quote requests by email. The goal handed to the agent: prepare a reply for every new request and leave it as a draft for the office manager.

A request comes in. The agent reads it and pulls out the address and the service needed. It checks the address against the service-area list and finds the address is covered. It checks the calendar for the next open estimate slots. It writes a reply offering two times and saves it as a draft. Then it stops.

The next request has no address. The agent does not guess. It drafts a reply asking for one, marks the draft as incomplete, and stops.

Two requests, two different paths, and nobody wrote an if-then rule for the missing address. That is the loop doing its job.

Notice what the agent never did. It never sent anything. A person reads every draft before it leaves the building. For a first agent, that checkpoint is not optional.

Where the Hype Comes In

The definition matters because the label is being applied generously.

On June 25, 2025, the research firm Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, pointing to escalating costs, unclear business value, or inadequate risk controls. The same release described what Gartner calls agent washing: existing assistants, chatbots, and automation tools rebranded as agents without much real agent capability behind them. Gartner estimated that only about 130 of the thousands of vendors claiming agentic AI were the real thing.

Two cautions on reading that. It is a forecast, not a result, and as of September 30, 2026, the deadline is still more than a year out. It also describes enterprise projects, not a small business trying one narrow task. The useful part for a smaller company is the list of reasons. Cost, value, and controls all get decided before a project starts, which means they are within your reach.

The people who build the models give similar advice. Anthropic's guide recommends finding the simplest solution that works and adding complexity only when it is needed, and it says directly that this can mean not building an agent at all. It also notes that agentic systems often trade speed and cost for better task performance. A loop that takes twelve steps costs more and runs longer than a script that takes three. Sometimes the trade is worth it. Often a checklist or a fixed automation does the job.

Four Questions Before Anyone Sells You One

When a product page says "agent," these questions get you to the truth quickly. They are also worth putting to an AI consultant before you hire one.

1. →What tools can it use? Ask for the list. If the answer is that it chats with your customers, it is a chatbot.
2. →Who decides the next step? If every path was drawn in advance by a person, it is an automation. That may be exactly what you need, at an automation's price.
3. →Where does it stop and ask a human? A good answer names specific checkpoints. "It doesn't need to" is not a good answer.
4. →Can I see what it did? Every step an agent takes should leave a record you can read afterward. If you cannot review its work, do not hand it work that matters.

Your Homework Before Part 2

You do not need software yet. You need one candidate task. The people who get the most out of AI tools tend to start the same way, with one narrow job rather than a platform. Look at your own week and write down a task that meets these conditions:

• →It repeats at least weekly.
• →It has a clear finish line. You could tell a new hire what "done" looks like.
• →Everything it needs is already digital: an inbox, a spreadsheet, a shared folder, a calendar.
• →The output can sit as a draft for a person to review before it goes anywhere.
• →A mistake caught at review costs minutes, not money or reputation.

Tasks that involve regulated communications, payments, or anything a customer would see without review are not first-agent material. Leave those alone for now, and bring in your compliance or legal advisor before you ever revisit them. If you run a regulated business, our getting-started guide for collection agencies shows what a compliance-reviewed first pilot looks like.

In Part 2, we take that one task and write it up the way you would brief a new employee: the goal, the tools, the limits, and the checkpoint. That document turns out to be most of the work of building an agent, and it takes no code.

If you would rather talk through where AI fits in your business before you build anything, see how we work with clients.

Sources

• →Anthropic, "Building Effective Agents" (December 19, 2024)
• →Simon Willison, "I think 'agent' may finally have a widely enough agreed upon definition to be useful jargon now" (September 18, 2025)
• →Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (June 25, 2025)

Sources cited above were checked on September 30, 2026. The quote-request example is hypothetical and for illustration only.