From answering to doing
Most AI you've used answers: you ask, it replies, you decide what to do next. An agent is different. You give it a goal, such as “sort out this customer's delivery problem”, and it works out the steps: look up the order, check the courier status, read the refund policy, draft a reply, and maybe issue a refund.
That's the whole idea. Five ingredients make something agentic:
- A goal rather than a single question.
- Tools it can use: search, a database, email, a calendar, another app.
- Memory of what has happened so far.
- Several steps taken on its own, in an order it decides.
- Checks and hand-offs: moments where it verifies its work or passes to a person.
Why this is a design problem, not only an engineering one
When software can act, the most important questions stop being about layout. They become questions about trust. How much should it do before asking? What does it show while it works? What if it's 70% sure? Who notices when it's wrong, and how does the person undo it?
Those are design decisions. Engineers can make an agent capable; designers decide whether people can live with it.
A simple autonomy ladder
One useful way to design an agent is to decide, task by task, which rung of the ladder it sits on. Low-risk, reversible tasks can go higher. Anything involving money, people's data or things that can't be undone usually stays lower.