Agents
AI Agent
A tool waits for the next command. An agent is given a goal and decides for itself which steps lead there: it can look things up, calculate, call another system, check its intermediate result, and change course. That shifts what operating means: you no longer give instructions for individual steps but set goals, limits, and exceptions. Two properties follow directly. The agent copes with cases nobody anticipated, which is the benefit. And in unanticipated cases it does things nobody anticipated, which is the risk, and it can't be tested away, only limited and observed. That is why the important decisions in agentic systems are organizational: what may it see, what may it do without asking, and how does someone notice that something is going wrong?
An example
A tool waits for the next command. An agent is given a goal, “Resolve this complaint,” and decides on the steps itself: look up the order, check proof of delivery, read the policy, calculate the credit, draft the reply email. It may take five steps or fifteen, in different order, and on two identical cases in a row it may choose two different paths. Therein lies both the benefit and the problem.
A closer look
An agent consists of a model, a loop, and tools. The model proposes a next step, the loop executes it, the result comes back into the context, and this repeats until a termination criterion. Everything else is elaboration.
Why it matters
As soon as software plans steps itself, leadership shifts. You no longer issue commands but take responsibility for goals, limits, and exceptions, and to do that you have to decide beforehand what the agent may see, what it may do without asking, and how a human notices that something is going wrong. These three decisions are the real effort in agentic systems. They are organizational, not technical, and they can't be retrofitted without reopening the workflow.
Commonly misunderstood
Errors accumulate across steps. At a 95 percent hit rate per step, a twenty-step workflow is wrong in roughly every third run. Short chains with checkpoints are therefore more robust than long autonomous ones.
Where this term does its work in the magazine.
2 articles use the term or take up its principle.