AI agent

Anthropic's engineering team defines agents as systems where the model directs its own process and tool use. At work, your part shifts from doing each step to briefing the agent and checking what comes back.

How it differs from an AI workflow

In an AI workflow, a person or a piece of code fixes the order of steps in advance, and the model handles one step at a time. An agent decides the order itself, based on what each tool call returns. That makes agents useful when you can't predict how many steps a task will take, and harder to debug when something goes wrong.

Anthropic's advice is to start with the simplest setup that works and add agent autonomy only when the task needs it.

Example

An office manager asks a coding agent to sort six months of downloads into project folders. She works on a copy and tells it to leave the Invoices folder alone. The agent lists the files, proposes a folder plan, moves files in batches and reports the ones it couldn't place. She watches the first few moves and pauses it once, when it is about to rename a file inside Invoices.

Common mistake

The usual first mistake is an open-ended job like "clean up my whole drive", where nobody can say when it is done or whether it went right. Pick a first task that is bounded, easy to check and easy to undo. Then read the agent's plan and let it run, instead of correcting every step.