An agent is a language model given a goal, a set of tools it may use, and permission to take several steps on its own before coming back. That last part is what separates it from a chat window: it decides the order of its own actions rather than waiting for your next message.
Four Parts Every Agent Has
Look for four components. Instructions state the goal, the method and the limits. Tools are the things it can actually do: read a mailbox, search a database, write a row, send a draft. Memory or context is the material it works from. The loop is the cycle of deciding, acting, reading the result and deciding again.
The Loop Is the Whole Point
In a fixed workflow you specify every step. An agent is given an objective and works out the sequence, which is valuable when the path varies by case: a request might need one lookup or four, depending on what the first lookup returns. That flexibility is also the source of the risk, because the sequence is not fully predictable.
What That Means for Operators
Because the path varies, you manage an agent through its goal, its tools and its boundaries rather than by scripting it. Give it the narrowest set of tools that completes the job, make its output reviewable, and treat the instruction set as a living document you tighten as you see real behavior.
Action Step
List three tasks you do weekly and mark which involve several steps whose order changes case by case. For each marked task, write down the tools or systems a person needs to complete it.
This lesson is educational and reflects general practice. Results depend on your situation, market and effort; adapt what you learn to your own context.