The short version
- A chatbot answers. An agent does. You ask a chatbot a question. You give an agent a job.
- Four parts: a goal, a model that plans, tools it is allowed to use, and a check on the result.
- The loop is the key. An agent acts, looks at what happened and decides the next step, again and again, until the job is done or it gets stuck.
- It can still make mistakes, including doing things nobody asked for. Good agents ask a person before anything important goes out.
A definition without the jargon
The big technology companies broadly agree on what an agent is. Google describes AI agents as “software systems that use AI to pursue goals and complete tasks on behalf of users” (Google Cloud). Amazon puts it this way: “Humans set goals, but an AI agent independently chooses the best actions it needs to perform to achieve those goals” (AWS).
Meta’s version is the most compact: “bots follow rules, assistants follow instructions, and agents execute toward goals” (Meta AI).
The four parts
1. A goal. Something a person hands it, such as “find three venues for the club dinner under €2,000 and email me a comparison.”
2. A model that plans. Most agents use a large language model, the same kind of AI behind ChatGPT, Claude or Gemini. IBM notes that this is why agents are often called “LLM agents” (IBM). The model breaks the goal into steps.
3. Tools. These are the things it has permission to use: a web search, your calendar, a spreadsheet, an email draft. Without tools, a model can only talk.
4. A check. The agent looks at what it produced, fixes obvious gaps, and either finishes or hands the work back to a person.
If any part is missing, you have something useful but not an agent. A model with no tools is a chatbot. Tools with no model making choices is ordinary automation.
A worked example
Say you ask an agent to tidy your inbox. Given permission, it might sort the messages, move dates and deadlines onto your calendar, flag the few that look important and draft replies for you to approve (Meta AI). One instruction, several steps, and you decide what is sent.
The important part is the middle. A fixed automation would do exactly the same thing every time. An agent can decide that one email needs a reply, another needs nothing and a third should go to a colleague.
Where it goes wrong
The New York Times’ new AI glossary makes the risk plain: agents “can make errors and may perform actions that were never requested,” and letting them send messages or use a credit card “can lead to complications” (New York Times).
Three common-sense rules help:
- Give it the least access it needs. Read-only is safer than read-and-write.
- Keep a person in the loop for anything that leaves the building: payments, emails to customers, public posts.
- Ask to see its working. A good agent can show which steps it took and why.
Why it matters
For most people, agents will arrive quietly, inside email, office software and phones, rather than as a separate product. The useful question is not “is this an agent?” but “what is it allowed to do on my behalf, and who checks it?”
Join the discussion
Would you let an AI agent send emails or make payments for you? Where would you draw the line?