Technology LearningThe word agent is thrown around so freely that it has started to mean nothing. Every product is suddenly agentic. Underneath the marketing there is a genuine, useful distinction, and once you see it you can tell a real agent from a chatbot with a costume on.
The line that actually matters
A plain chatbot answers. You ask, it replies, and the exchange ends there. It cannot do anything in the world beyond producing text.
An agent acts, and decides. Give it a goal and it can take steps toward that goal, use tools to affect things outside the conversation, look at the result, and choose what to do next, repeating until the job is done. The defining move is that loop: act, observe, decide again. A chatbot has no loop. An agent runs on one.
The two ingredients
Strip away the hype and an agent needs two things a chatbot lacks.
- Tools. A tool is anything the model can call to get information or make something happen: search the web, read a file, query a database, send a message, run a calculation. Tools are how an agent reaches past its own text and touches the real world.
- A loop with judgement. The agent does not run a fixed script. After each step it looks at what came back and decides the next move. That decision-making inside the loop is what makes it an agent rather than an automation.
Take away the tools and it is a chatbot. Take away the loop and it is a single tool call. You need both.
A quick example
Ask a chatbot "is this invoice overdue?" and it can only reason about what you pasted in. An agent given the same goal can look up the invoice, check today's date, compare the two, and tell you, and then, if allowed, send the reminder. Same question, but one can only talk about it and the other can go and find out and act.
Why the loop needs limits
The same loop that makes an agent powerful is what makes it need guardrails. Something that can take actions can take wrong ones, and something that decides its own next step can decide badly. Real agents are built with limits: what tools they may use, which actions need a human to approve, and when to stop. Designing those limits is not an afterthought, it is most of the actual work.
It is a spectrum, not a badge
Do not treat agent as a label a thing either has or lacks. A system can be a little agentic, calling one tool and stopping, or deeply agentic, running long chains of decisions across many tools. Most useful things sit somewhere in between, and the right amount of autonomy is the amount the task needs and no more.
Where to go next
Understanding the act-observe-decide loop is the foundation for building anything genuinely agentic, and it saves you from both under- and over-engineering. What an Agent Actually Is sets the concept out plainly, Tools covers the piece that lets an agent reach the world, and The Five Patterns shows the common shapes real agents take.