The same word is being used for two different things. On one side sit the chat interfaces you ask a question and get an answer from. On the other sit the agents that are said to finish work on their own. Both are very likely calling the same model, often the same version from the same provider.
The model is not what makes the difference. What gets built around it is. Four behaviours tell them apart.
Calling tools
A chatbot produces text and stops there. Part of what an agent produces does not stay text. It turns into an action.
On the technical side the model is still writing. But what it writes is a structured request that means "run this tool with these parameters", a system outside the model runs it and hands the result back. The model looks at the result and decides on the next step.
A chatbot tells you "update that record". An agent updates it.
Working in several steps
A chatbot starts and ends in one turn. It gives whatever the answer is and you open the next step.
An agent breaks the work up itself. It decides what to do first, runs one step, looks at the outcome and changes the plan if it has to. When it runs into something unexpected halfway, it backs up and tries another route.
There is a side of this you do not see from outside. Only a small part of the text produced in an agent run ever reaches the user. The rest is tool calls, intermediate results, notes the agent keeps for itself and descriptions of sub-tasks. Behind the single paragraph you read sit dozens of steps, and that invisible part is where the cost comes from.
Checking its own output
A chatbot leaves correctness to you. You read the answer, ask again if you do not like it, and you are the one running the correction loop.
In an agent that loop is built in. An acceptance criterion is defined up front, the output is tested against it and retried when it fails. Running the test, reading the result and fixing it all happen inside one run.
Not every agent does this. But it gets hard to call a system an agent when it never does, because a setup that does not check its own work is really a chatbot with some tools wedged in.
Deciding to stop
This fourth difference is the least discussed and the hardest to build.
A chatbot has no stopping problem. It gives the answer and closes the turn, and opening the next one is your job.
An agent has no natural end like that. It has to decide for itself that the work is done. When that decision goes wrong the damage comes from two directions. Stop too early and the work is half finished, never stop and it spins on the same step and runs up the bill.
Teams running agents spend most of their effort right here, not on producing output. Well-built systems do not leave the stopping decision to the model alone, they put a ceiling on it from outside. That is what step limits, spend limits and time limits are for.
The grey zone in between
Most products do not carry all four behaviours. There are systems that can call tools but do not choose the order of their own steps, and most of the "agent" labels on the market are describing exactly that.
Anthropic's engineering post draws the line clearly. In workflows, the model and its tools are run along "predefined code paths", while in agents the model "dynamically directs its own processes and tool usage" (Building Effective Agents).
That is where the distinction comes down to one question. Does code set the order of the steps, or does the model?
This is not a ranking of quality. For work whose order is known in advance, a code-directed setup runs cheaper, faster and more predictably.
Why the distinction matters to a buyer
The cost of a mistake is not the same on both sides.
When a chatbot answers wrongly you read a wrong text. If you notice, you fix it, and if you do not, you carry on with wrong information.
When an agent decides wrongly an action happens. A record is updated, a message goes out, a file gets written. It is not always possible to undo.
So the questions to ask when buying an agent are different from the ones for a chatbot. You are not asking about the quality of the answer. You are asking about the system's authority and its limits.
Three questions for the vendor
Which tools can the system call on its own decision, and who sets that list?
How many steps can one job take at most, what makes it stop, and what happens when it does not?
Can I see afterwards which steps were taken in a finished run?
If you cannot get a clear answer to all three, what you are looking at is most likely a chatbot that uses tools. That does not make it a bad product, but it is a different category and the budget should be judged accordingly.
Sources
Anthropic, Building Effective Agents