If you have read ten definitions of “AI agent” and come away more confused, you are not alone. Vendors use the word for everything from a chatbot with a new coat of paint to a fully autonomous software engineer. This post gives you one working definition and the mental model to spot the difference.
The one-sentence definition
An AI agent is a language model that runs in a loop: it looks at a goal, decides on a next action, carries that action out with a tool, reads the result, and repeats until it decides the goal is done.
That loop is the whole trick. A chatbot gets one message and returns one answer. An agent gets a goal and returns a finished piece of work, with many model calls and tool calls in between.
The four parts every agent has
Strip away the branding and every agent has the same four parts.
1. A goal
Something a human wants done: “find three suppliers under $2 per unit and draft an email to each”, or “fix the failing test in this repository”.
2. Tools
Functions the model is allowed to call. Web search, a code runner, a database query, a browser, an email sender. The tool list defines what the agent can touch, so it is also the main safety control.
3. Memory
At minimum, the running transcript of what it has done so far. More advanced agents also store notes between runs so they stop repeating mistakes.
4. A stopping rule
How the agent knows it is done. Good agents verify their work (run the tests, re-read the output) before stopping. Bad agents stop when they feel like it.
Where agents already work well
- Coding tasks with tests. The test suite is a built-in stopping rule, which is why coding agents matured first.
- Research and summarisation. Search, read, compare, write. Low risk if a human reads the result.
- Data cleanup. Normalising spreadsheets, deduplicating records, mapping fields between systems.
- Support triage. Reading tickets, tagging them, drafting a first reply for a human to approve.
Where they still fail
- Long, open-ended projects. Errors compound across hundreds of steps. Humans still need to review checkpoints.
- Anything irreversible without a check. Sending money, deleting data, emailing customers. Put a human approval step in front of these.
- Tasks where “done” is fuzzy. If you cannot describe how you would check the work, the agent cannot either.
How to tell if a product is really an agent
Ask three questions. Does it take a goal rather than a message? Does it call tools and read the results? Does it decide when to stop? If the answer to all three is yes, it is an agent. If it is a chat window with a bigger prompt, it is a chatbot, and that is fine too, just call it what it is.
What to read next
Our tutorial on building your first agent walks through each of these parts with code, and the Tools page lists the frameworks we have tested.
Questions, answered
What is the difference between an AI agent and a chatbot?
A chatbot answers one message at a time. An agent takes a goal, breaks it into steps, calls tools such as search, code or APIs, checks the results and keeps going until the goal is met or it gives up.
Do AI agents work reliably in 2026?
For bounded tasks with clear success checks, yes: coding, data cleanup, research summaries and support triage. For open-ended multi-day work, they still need a human reviewing checkpoints.
Do I need to code to use an AI agent?
No. Tools like n8n, Zapier Agents and Claude in Chrome let you run agents without writing code. Building your own custom agent still needs some programming.