Budget first, then run
Why agents cost more
Each step an agent takes sends text to the model and gets text back, and the conversation grows as it goes. A task of fifty steps can use as much as hundreds of ordinary chat messages. On subscription plans this shows up as usage limits and slow-downs; on pay-as-you-go API keys it shows up as money, charged per token.
Runaway loops
Sometimes an agent gets stuck: retrying a failing step, re-reading the same page, or 'improving' a result forever. Without a cap it keeps spending. Tell it how many attempts it may make, set a maximum number of steps or a time limit where the tool allows, and ask it to stop and report if it hasn't finished.
Set hard limits
If you use API keys, for example with a self-hosted agent, set a monthly spending cap in the provider's dashboard and a low per-key limit. Use a separate key per agent so you can see and cut off each one. Check the usage page weekly at first. Cheaper models are often good enough for routine agent steps.
Know the off switch
Before you start anything long-running, find out how to stop it: the stop button, how to cancel a scheduled task, how to revoke an app connection, how to delete an API key. Practise once. When something goes wrong, the minutes it takes to find the off switch are when most of the damage happens.
If an agent ran in a loop all night, would you find out from a notification, a usage warning or your card statement?