When Should an AI Agent Stop and Ask for Help?

An agent that retries every failed tool call can turn a small glitch into a larger problem. But escalating every hiccup spends human attention too—and that resource is not unlimited.

For a workflow that can affect real records or customer service, where should the line be between retrying, pausing, and asking for approval? I’d want the answer to account for whether the action can be safely reversed, whether permissions or inputs changed, and who owns recovery if the agent stops midway.

What signals would you use to set that boundary without making the workflow either reckless or needlessly dependent on constant supervision?