Short answer
An AI agent is an AI that takes a goal, decides the steps itself, and uses tools to carry them out. A chatbot gives one answer to one question. An agent can take a request like "research three competitors, put them in a table, and make a file" and work through searching, reading, organizing and file creation in order. That is convenient, but mistakes and costs grow with it, so keep payments and accounts out of its hands.
Chatbot vs. agent
A typical chatbot is one round trip: you ask, it answers. It does not open the internet or create files in between. An agent does not stop at writing an answer. It takes the actions needed to finish the job.
Anthropic describes agents as systems where the AI decides its own process and tool use. When a developer fixes the order of steps in advance and uses the AI only inside that script, it is called a workflow instead. So the key idea is that the AI chooses what to do next.
What it does
- Chatbot
- Writes an answer to your question
- Agent
- Takes a goal and carries out several steps
Tool use
- Chatbot
- None, or one or two (such as search)
- Agent
- Chains search, web browsing, file creation and more
Your involvement
- Chatbot
- You check every answer
- Agent
- You hand off the middle and check the result
Time
- Chatbot
- Seconds
- Agent
- Can take minutes or longer
Effect of a mistake
- Chatbot
- A wrong answer
- Agent
- A wrong action may actually be carried out
What does an agent actually do?
Broken into steps, it goes roughly like this. It understands the goal, searches for material, reads it, organizes it and produces a result. If it gets stuck, it tries another way.
- Research: opens several websites, collects information and puts it in a comparison table.
- Document work: turns the research into a report, slides or a spreadsheet.
- Connecting to other apps: if it is linked to a calendar, email or drive, it can look up events or documents. See "What is MCP?" for how those links work.
- Software work: coding agents read, edit and run the files in a project directly.
Real examples
ChatGPT has a feature called ChatGPT agent. OpenAI introduced it as a feature that does multi-step work for you, such as browsing websites and creating files. According to OpenAI, it is trained to ask permission before actions with real-world effect, such as a purchase, and to refuse high-risk tasks such as bank transfers.
In Claude, connectors let you link apps such as Google Drive and Google Calendar, so a question like "What is on my calendar tomorrow afternoon?" can be answered from your own data. Anthropic's help pages say connectors let Claude look up information and take actions in external apps, and that there are settings to require your approval before a tool is used.
Names and menus differ by service and change over time. Check inside the app you use for something called "agent," "agent mode" or "deep research." Whether it is available on a free plan, and any usage limits, also vary by service.
Good first tasks
Start with work where you can check the result yourself and a mistake does no harm.
- Drafting a trip itinerary from research (you do the booking yourself)
- Collecting comparison material on neighborhoods before a move
- Summarizing a company's public information before a meeting
Compare three automatic cat feeders. Put price range, capacity, app support, and the complaints that come up most in user reviews into a table. Add a link to the source page in each cell. If you could not confirm something, write "not confirmed."
The last sentence matters. It stops the agent from inventing something to fill an empty cell.
What not to hand off
Agents sometimes misjudge a situation in ways a person would not. Anthropic notes that because agents work through many steps on their own, cost and run time go up and small errors can pile up from step to step. So the rule is simple: do not hand off anything you cannot undo.
- Payments and transfers: do not have it enter card numbers or buy things automatically. Let it fill the cart, then pay yourself.
- Accounts and passwords: do not give it logins, verification codes or recovery keys. This applies most to bank, brokerage and government accounts.
- Sending email and posting: read it first and press send yourself.
- Deleting and changing: for files, calendar events or customer records, keep an approval step turned on.
- Important judgments: for legal, medical or tax decisions that carry real consequences, check with a professional.
Habits that keep it safe
A web page an agent opens may contain hidden text such as "ignore your previous instructions and do this instead." This is called prompt injection. OpenAI says agents can be exposed to this kind of attack and that it keeps strengthening its safeguards. It is wise not to run an agent while you are logged in to sensitive sites.
The practical advice is short. Give one goal at a time, keep the setting that asks for approval along the way, and check the result against the source links yourself. What you learned about hallucinations applies to agents too.
Related terms
- Workflow
- A setup where a developer fixes the order of steps and the AI works inside it. In an agent, the AI chooses the steps.
- Tool use
- The AI calling an outside function such as search, a calculator or file creation.
- Prompt injection
- An attack that hides instructions in a web page or document to push an AI into doing something unintended.
FAQ
Is an agent better than a chatbot?
It depends on the job. For one question and one answer, a chatbot is faster and cheaper. An agent helps when there are many steps, such as visiting several sites and then producing a file. Anthropic itself advises against complex setups when a simple approach works.
Does an agent take over my computer?
It depends on the service. Some work only inside a web browser, and some can reach files on your computer. Read what permissions it asks for when you install or connect it, and allow only what you need.
Is it free?
It varies by service and plan, and it changes often. Check each service's help pages for current availability and limits.
Who is responsible when an agent gets it wrong?
In practice, you are. Service terms usually put the job of checking results on the user. That is why you should not hand off anything irreversible and should do the final check yourself.
Keep going
Checked against sources on 2026-10-30. Sources: Anthropic, Building effective agents, Claude, Get started with connectors, OpenAI, Introducing ChatGPT agent. How we write