A model given tools and allowed to take steps.
An agent wraps a model in a loop: decide an action, call a tool — search, code execution, an API — read the result, decide again, until the task is done.
Capability comes from the tools as much as the model. The failure modes are compounding: one wrong step early produces confident nonsense several steps later.
Serious deployments therefore constrain agents heavily — limited tools, step budgets, and a human check before anything irreversible.