A system where a language model does not only answer, but also calls tools — searching a database, creating a ticket, sending a message — in a loop until a task is finished.
Agents can take over multi-step routine work that previously needed a person. They also act with real consequences, so the design question is not capability but what the system is allowed to do unsupervised.
We give agents narrow, explicitly defined tools, require confirmation before anything irreversible, log every action for audit, and cap loops and spend so a misunderstanding cannot run up a bill or a mess.