Our agent made 21 unwanted calls and cost $133 before we added one hard cap. Here is the exact setup that stopped it.
retained text
Story, no pitch - the setup is at the end and the tool we built is in my profile if anyone wants it. What happened: our outbound agent worked fine for weeks. Then one instruction edit - literally one line about being proactive - and overnight it started offering calls to leads who never asked for one. 21 calls in a day, $133 in telephony + LLM spend, and two confused prospects. The uncomfortable lesson: the prompt did not drift gradually. It jumped, all at once, across every conversation, the moment the instructions changed. If you only measure what the agent sent, you find out from the leads. The fix was not a better prompt. It was three gates outside the prompt: 1. Hard dollar cap per action. The agent literally cannot spend more than X on one call/email/LLM run. The cap does not care what the instructions say. 2. Velocity limit. Max 10 outbound actions per hour. A retry loop or a bad edit can't spam - it just stalls. 3. Human approval queue for anything unusual. If the action type is new, or the spend is 2x the median, it waits for a person. Since then: zero incidents across thousands of actions. The cap never fires on normal traffic - it only exists for the day someone edits instructions at 3am. Ask me anything about the setup - happy to share details, including what did NOT work (trial budgets, prompt warnings to the agent, and trusting the model to self-limit - all failed before this).
document fetched 2026-11-10 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xrme6d/our_agent_made_21_unwanted_calls_and_cost_133