AI agents: what they are, and what it takes to run one

What is an AI agent?

An AI agent is a system that is given a goal and a set of tools, and works toward that goal by taking actions and reacting to the results. It perceives some state, decides on a step, executes it, observes what happened, and repeats until the goal is met or it stops.

The word doing the work in that sentence is acts. A language model on its own produces text. An agent is that model wrapped in a loop, handed tools, and pointed at an outcome — which is why the interesting questions about agents are not about how clever the model is, but about what it is allowed to touch.

Agent, chatbot, assistant: the distinction that matters

• A chatbot answers. It replies from a script or a set of documents, and cannot change anything. Ask it to update a record and it will describe how you might.
• An assistant helps you work. It is excellent inside a window you have open, on a task you are actively doing. Close the window and nothing continues.
• An agent is given a job. It can run when nobody is watching, start because something happened rather than because you asked, and change things in systems outside the conversation.

The practical test: can it do something you would have to undo? If yes, it is an agent, and it needs the governance the rest of this page describes. If no, it is a chatbot with better writing, and most of these concerns do not apply. The longer version is in AI agent platform vs chatbot.

The four parts every agent has

Whatever the vendor, the architecture underneath is consistent:

• Reasoning. A model interprets the goal and decides what to do next. This is the part that improves when the industry ships a better model, and the part people wrongly assume is the whole system.
• Planning. Breaking a goal into steps, and adjusting when a step fails. A one-shot answer is not an agent; the loop is.
• Memory. Short-term, so it can follow a conversation; long-term, so it does not start from nothing every morning. Memory is also how an agent learns your procedures — see how AI agents remember.
• Tools. The connections that let it act outside the chat: an API, a database, a browser, a calendar, your CRM. An agent with no tools is an assistant.

A fifth part belongs on that list for any agent doing real work, and vendors rarely include it: a boundary. What it may do unsupervised, what waits for a person, and how you reconstruct afterwards what it did.

What agents are genuinely good at today

Being specific is more useful than being enthusiastic. The categories that work reliably now:

• Answering from a known body of knowledge — your documents, your price list, your policies, your inventory — across whatever channel the question arrives on.
• Moving information between systems that were never integrated: an enquiry becomes a CRM record becomes a calendar entry becomes a note.
• Recurring work on a schedule — monitoring, digests, reports, first drafts waiting for you in the morning.
• Research and synthesis — reading a lot of sources and producing something short and structured.
• Triage — sorting what arrives into what matters, what can be answered automatically, and what needs a person.

And where they are still weak: anything requiring judgement about people, anything where being confidently wrong is expensive and unverifiable, and anything with no clear signal of whether the job was done correctly.

What running one actually requires

This is the part the category page usually skips. The model is the easy bit; five other things decide whether a deployment survives its third month:

• Somewhere to run that stays up. An agent on a laptop stops when the laptop sleeps, which defeats most of the point — see always-on AI agents.
• Connections that keep working. Tokens expire, OAuth grants lapse, endpoints move. Reconnecting is ongoing work, not setup.
• Grounding. An agent answering from your documents states your facts; one answering from general knowledge invents plausible ones. This is configuration, not model quality.
• An approval boundary and an audit trail. Decided deliberately, per action, before the first real task.
• Spending controls. An agent that runs on a schedule can run up a bill; caps should be hard rather than advisory.

One agent or several?

A single agent with every tool and a long prompt works until the instructions start contradicting each other and you can no longer tell which part caused a bad answer. Splitting the work into agents with distinct jobs fixes that — the same way it does with people. That introduces orchestration, which is its own subject: multi-agent AI platforms.

Do agents learn?

Yes, but not in the way the word suggests. The model's weights do not change. What changes is everything around it: what it remembers, the procedures it has been taught, the skills it has written down. That distinction decides what you can reasonably expect, and it is worth understanding before you evaluate any "self-improving" claim — self-improving AI agents covers it.

Where to start

One workflow. Pick the one that costs you most today — in hours, in errors, in enquiries nobody answered — and give that to an agent with a narrow boundary. Widen from evidence rather than optimism. Everything else on this site is a spoke off that idea: what AI changes in a business, how to choose a platform, and the tutorials if you would rather see one built than read about it.