An AI agent is a language model given tools and a loop: it reads a goal, picks a tool, looks at the result and goes again until a stop rule says it is done. A chatbot answers and waits for you; an automation runs fixed steps in a fixed order. The agent decides the steps as it goes, which is why the right first agent comes from a task you already do by hand and know the finish line for.
Agent vs chatbot vs automation
|
Chatbot |
Automation |
AI agent |
| What it does with input |
Answers it |
Runs a fixed sequence on it |
Works toward a goal with it |
| Who decides the steps |
You, one message at a time |
You, once, when you build it |
The model, at run time |
| Tools |
None, or a search box |
Fixed connectors, in order |
A set it chooses from |
| When it stops |
After each reply |
At the end of the sequence |
When the stop rule is met |
| How it fails |
Wrong answer |
Breaks when the input changes shape |
Loops, or picks the wrong tool |
| Example |
Ask ChatGPT for a subject line |
New form entry creates a CRM row |
Read the entry, research the company, draft the reply, file both |
An automation is the right choice whenever the steps never change. Build an agent when the steps depend on what comes back. If you cannot say in advance whether step two is "search the site" or "email the client", you need something that decides.
The parts of an agent
Five parts, and you will spend your time on the last two.
- The model. It reads the goal and the tool results and chooses the next move. Any current model with tool calling works.
- The tools. Each one is a function the model can call: search, read a file, query a database, send an email, run code. The MCP servers directory lists connectors you plug in instead of writing.
- Memory. What the model can see right now (the context) and what it can look up (a file, a database, a search index, a note it wrote earlier). Long jobs need the second kind, because the first fills up.
- A goal. One sentence with the finished state, plus the rules. This lives in the system prompt, and the system prompt generator guide gives you a structure for it.
- A stop rule. The condition that ends the loop, and a hard cap on steps for when the condition never arrives. An agent without a cap is a bill.
Skills sit alongside these. A skill is a written procedure the agent loads when the task matches, so a job it has done well once gets done the same way again. The skills directory has ready ones.
Your first agent in five steps
- Pick a task you did by hand this week. It should take 15 to 60 minutes, have an obvious done state, and touch at most two tools. Weekly competitor price check. Inbox triage into four labels. Meeting notes into tickets. Receipts filed by client.
- Write the run sheet. What would you tell a temp on day one? Inputs, the steps as you do them, what a good result looks like, and what to do when stuck.
- Turn the run sheet into the system prompt. Goal at the top, rules next, the steps, then the report format. The system prompt generator guide linked above is built for this step.
- Give it the two tools and nothing else, then run it on one real input while you watch. Every step it takes is a line you will either keep or fix in the prompt.
- Add the stop rule and the step cap, run five more inputs, and compare its output with what you would have produced. Fix the prompt before you touch the tools; early failures live in the instruction far more often than in the tool.
Where to build one
Claude Code, for agents that live in a terminal and work with files, repos, shell commands and MCP tools. You write the goal in plain text, point it at the tools, and it runs the loop for you. Skills plug straight in, and the Claude Skills Mini Course walks through building one from scratch.
n8n, for agents that connect apps on a trigger: a new email, a form entry, a schedule, a webhook. It is visual, the connectors for the common SaaS apps already exist, and the agent node sits inside a normal workflow. If you would rather start from working workflows than a blank canvas, the n8n automations bundle gives you a set to open and adapt.
The choice matters less than the run sheet. The same sheet works in any of them, because the loop is the same loop.
What breaks
- The loop never ends. There was no stop rule, or the done condition can never be true. Put a step cap on every agent, on day one.
- It reads an error as success. A tool returns "no results" and the model treats that as the answer. Make the tool return an explicit failure, and put "if a tool fails, say so and stop" in the rules.
- The context fills up. Twenty steps in, it forgets the goal. Have it write a running summary to a file and reread it each step.
- Too many tools. Give it twelve and it will call the wrong one. Start with two.
- The goal lived in your head. The prompt said "clean up the leads" and you meant "dedupe by email, keep the newest". Write the finished state, in detail.
- An irreversible action with no gate. Sending, deleting, paying, publishing. Every one of those needs a confirm step until you have watched fifty clean runs.
What is inside the AI agents guide
The download is free, and it stays with you:
- AI Agents Mini-Course
- How-to Guide and Tips
- Its one-line description: "Automate Your Business and Life with Powerful AI Agents"
- Lifetime access and updates
