An agent-tool loop is the cycle in which a model observes the state, decides on an action, calls a tool, reads the result, then starts again until it reaches the goal. It gives the model autonomy, hence the need for strict limits.
Example
observe -> decide -> call a tool -> read the result -> repeat