Watch one agent run for an hour and count the retries. A tool call fails, the agent tries again, same input, same failure, three, four, five times. Each retry is a full model call. The bill is not the model price, it is the loop. The fix is boring and it works: a retry budget. Two tries on the same input, then stop and escalate. Not five, not ten, two. And log the failure with the exact input that broke it, so you fix the cause instead of paying for the same failure again tomorrow. An agent that retries forever is not persistent. It is a leak with a chat interface. How many retries do you allow before it stops?