Autonomous agent

Autonomy is about how many decisions an agent makes on its own before a person checks in. A fully autonomous agent receives a goal and runs until it is done, approving its own actions within whatever limits it has been given.

Autonomy is a setting you choose, not a feature to maximize. It makes sense when the task is well defined, mistakes are cheap or reversible, and the agent works inside a sandbox with limited permissions. It makes less sense when actions are expensive, public, or hard to undo.

In practice, most useful autonomous agents combine freedom inside a safe boundary with hard checkpoints at the edges, such as requiring a human to merge a pull request or approve a payment.