What makes something an agent

A plain LLM call takes text and returns text. An agent is given tools — functions it can ask your code to run — and works in a loop: decide on an action, see the result, decide what to do next, until the task is done.

It helps to separate two designs:

• Workflows: your code decides the steps. "Classify the email, then look up the customer, then draft a reply." The LLM fills in each step. Predictable, testable, cheaper.
• Agents: the model decides the steps and how many to take. "Resolve this support ticket." Flexible, able to handle situations you didn't anticipate — but less predictable, and each step costs time and money.

Start with the simplest design that works. Many problems described as "we need an agent" are solved better by a single well-prompted call or a fixed workflow. Reach for an agent when the path genuinely can't be known in advance: open-ended research, debugging, tasks with many possible branches.

Agents are also only as good as the feedback they get. Tasks where the agent can check its own progress — tests that pass or fail, a search that returns results or doesn't — suit agents far better than tasks where nothing tells it when it is wrong.