It plans, acts, checks the result, and either carries on or comes back to you when it needs a decision.
How does an AI agent work?
Inside an agent, the model works out what to do next and the tools give it a way of doing it: a calendar, a messaging app, a file, a database, a web request. A layer of software around the model, the agent engine, runs the loop. The model picks a step, the engine carries it out, the result goes back to the model, and the cycle repeats until the job is finished or the agent has to ask you something.
Anthropic's guide Building effective agents describes agents in much the same way: systems where the model directs its own process and its own use of tools, instead of following a fixed script.
The everyday comparison is the gap between asking a colleague a question and handing them a job. A question gets you an answer and leaves the next step with you. A job comes back finished, or with one specific point for you to settle.
What can you hand over to an agent?
With a chat tool you still do the work around each answer: copying, pasting, sending, remembering to come back to it. An agent can take on that surrounding work, which makes the biggest difference on tasks that repeat and involve several steps, such as sorting what came in overnight, keeping a running list up to date or putting together a weekly summary.
Because an agent acts, the questions you ask before using one change too. How polished its writing is matters less than what it can reach, what it's allowed to do without asking, and how you see what it did afterwards.
An example: your own inbox
It's Monday morning and forty emails arrived over the weekend. Six of them ask when an order or a document will be ready.
With a chat tool, you open each email, paste it in, ask for a reply, tidy the draft and send it yourself, six times.
With an agent connected to your inbox, those six are picked out and sorted by urgency, and a reply to each is drafted in your usual tone using the delivery terms you've given it. The drafts wait for you. You read them, change a line in one and send them. One email asks about a contract the agent has no information on, so it flags that one for you instead of guessing. Nothing goes out until you approve it, which is what human in the loop means in practice.
How is an AI agent different from a chatbot?
A chatbot takes a message and returns a message, and it stops when you close the window. An agent keeps hold of a goal over time, uses tools, can run on a schedule (see scheduled task) and can carry on while you're doing something else. Most agents use a chat window as their front door, which is why the two get mixed up.
Common questions
Is an AI agent the same as an AI assistant?
Mostly, though both words are used loosely. "Assistant" usually suggests a conversational helper, while "agent" suggests something that takes actions towards a goal. Plenty of products sold as assistants are agents, and some sold as agents are only chatbots.
Can an agent keep working while I'm offline?
Yes, if it runs on a server rather than on your laptop. A hosted agent can run scheduled tasks overnight and have the result waiting in your chat app in the morning, while one running on your own computer only works when that machine is on and connected. The trade-offs are covered in hosted vs self-hosted AI agents.
What's the biggest risk with an AI agent?
An action you didn't intend, such as a message sent to the wrong person or a record changed or deleted. A poor draft costs you a minute to fix; a wrong action can cost far more. Approval before anything goes out, access limited to what the task needs, and a log of what the agent did keep that risk small. Our AI agent safety guide goes through each.