What Is an AI Agent? A Plain-English Guide With a Real One

An AI agent is software that is given a goal, works out the steps to reach it and uses tools to carry them out, checking its own progress along the way. A chatbot answers when you ask. An agent gets on with the job.

That definition gets stretched a lot, so here’s what it means in practice, the parts every agent has, and a real one we run every day.

Chatbot, Automation or Agent?

	What it does	Example
Chatbot	Answers one question at a time, when asked	“Summarise this document”
Automation	Follows fixed steps you designed in advance	“When a form is filled in, add a row to the sheet”
AI agent	Decides the steps towards a goal and uses tools to take them	“Answer every reply to our campaign and book the meetings”

The line that matters: an automation breaks when something unexpected happens, because nobody wrote a step for it. An agent reads the situation and chooses what to do next, inside the limits you set.

The Six Parts of Every Agent

1. A goal. One clear job with a clear idea of what good looks like. “Book meetings from replies” works. “Help with sales” doesn’t.
2. Instructions. The brief: who it’s working for, the tone, the rules, examples of good and bad. Most of the quality comes from here.
3. A model. The AI that does the reasoning and writing. Swappable, and usually the least interesting decision.
4. Tools. Access to the systems it needs: inbox, calendar, CRM, a search tool. This is where MCP comes in, a standard way of plugging agents into tools.
5. Context. Your own material: past messages, your tone of voice, your product, your customers.
6. Guard rails. What it must never do, what always goes to a person and who checks its work before it reaches a customer.

A Real One: The Agent That Answers Our Sales Replies

We run outbound campaigns over email and LinkedIn, and every reply lands with an agent. It:

• reads the reply and researches the person and their company;
• writes a personal answer aimed at that person;
• reads the live diary and offers two real meeting times;
• books the meeting when they confirm, and tells the team;
• follows up when someone says yes and then goes quiet;
• finds the right colleague when an out-of-office points to someone else.

The guard rails matter as much as the agent. A second agent reviews every draft before it sends and sends it back for fixes. Every automatic reply waits 5 minutes so a person can stop it. Pricing questions, objections it can’t answer and anyone asking about their data go to a person.

Over the 14 days to 28 Oct 2026, its median time to first reply was 6 minutes, including that 5-minute wait. 88% of first replies went out within the hour, and 8 in 10 messages went out with no human touching them. The live numbers and their definitions are on demandguy.com.

Where Agents Earn Their Keep First

Look for work that is repetitive, high volume, mostly reading and writing, and has a clear standard of good. Replies, account research, meeting notes, follow-ups and weekly reporting all qualify. Work that depends on a live human connection is harder; we still keep cold calling with people, because AI voices pause and it shows.

Build One Yourself

At Nerd Mode in Amsterdam you’ll build a first agent in an hour, connect it to real tools and see the reply agent above taken apart on stage. Register your interest for first access to tickets.