StudioOne prompt edit made our outbound agent 21 unwanted calls and cost $133 — three guardrails outside the prompt fixed it
Original title: Our agent made 21 unwanted calls and cost $133 before we added one hard cap. Here is the exact setup that stopped it.
The author's outbound agent suddenly offered calls to leads who never asked after a one-line instruction edit, making 21 calls in a day and costing $133 in telephony and LLM spend. The fix wasn't a better prompt but three gates outside it: a hard dollar cap per action, a velocity limit of max 10 outbound actions per hour, and a human approval queue for anything unusual or spending 2x the median. Since then there have been zero incidents across thousands of actions. The author notes trial budgets, prompt warnings, and trusting the model to self-limit all failed.
The author recounts an agent failure and shares three guardrails placed outside the prompt, offering transferable ideas for teams deploying outbound calling agents.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.