How an agent actually reasons
Why an agent does the work instead of describing it.
A chatbot answers. An agent acts. The whole difference is a loop that refuses to stop until the work is done.
Ask a chatbot a question and you receive a paragraph. Ask an agent to do something and it begins a cycle that does not end until the thing is finished. Or until it has good reason to come back and ask you for help.
Perceive, plan, act, observe
The agent observes the current state of the task. It forms a plan. It takes a single action through a tool: a search, a write, a calculation. Then it reads the result of that action and revises everything it believed a moment ago. Then it begins again.
Reasoning here is not a single answer. It is a sequence of small, checked decisions.
Why the loop matters
Each tool call turns language into a real effect: a row reconciled, a ticket routed, a vendor onboarded. And because the agent reads the outcome of every action before choosing the next, a wrong turn is no longer fatal. It is simply the input to the next decision.
This is the quiet revolution. Not a model that knows more, but a process that can notice it was wrong and try again, which is, after all, the oldest definition of intelligence we have.