Agent loops need budgets and stopping conditions

An agent that can decide to do one more thing will, given the chance, do one more thing indefinitely.

Without a limit on steps, time or spend, a loop that fails to converge keeps going: calling tools, spending money, making changes. The failure is rarely malicious, it is usually a task the model cannot complete and will not abandon. Budgets and stopping conditions are the difference between a stuck agent and an expensive incident.