How to stop AI agents from burning API credits in retry loops

Last month an autonomous research agent that I built ran overnight and burned through forty‑five dollars in API credits on a single broken URL. Of failing cleanly when a webpage returned a 403 error the AI agent kept rewriting its own search query and trying again every four seconds.

Adding a global timeout did not fix the problem because each individual request succeeded at the network level. What actually worked was adding a state‑tracking counter directly into the AI agent’s memory object. If the tool returns the error category twice in a row the AI agent forces a hard stop and logs the exact payload for human review.

I also started passing a remaining budget variable into the system prompt so the AI agent knows when to wrap up its work. Since making both changes our overnight batch runs have stayed under four dollars total without supervision.

Always cap consecutive tool retries, at the code level than relying on prompt instructions to stop a loop. How do you set up guardrails to catch AI agent loops before they drain your token budget?