One prompt edit made an agent burn $133 on 21 rogue calls — and the 3 gates that fixed it
Worth_Wealth_6811 · reddit · 2026-11-09
The author's outbound agent ran fine for weeks until one line asking it to be "more proactive" made it offer calls to leads who never asked: 21 calls overnight, $133 in telephony + LLM spend.
Key lesson: prompts don't drift gradually — they jump all at once across every conversation the moment instructions change, and if you only monitor what the agent sends, prospects are how you find out.
The fix wasn't a better prompt but three gates outside it:
- Hard dollar cap per action — the agent literally can't spend more than X per call/email/LLM run, regardless of instructions;
- Velocity limit — max 10 outbound actions per hour, so retry loops stall instead of spamming;
- Human approval queue for anything unusual (new action type, or spend 2x the median).
Zero incidents across thousands of actions since; trial budgets and in-prompt warnings to self-limit all failed before this.
More from coding & agent
- Agent researcher asks: how valuable is training beyond 256k context on Tinker? — cHHillee · 2026-11-09
- Capwise adds CLI + Doubao voice for auto screen-recording demos, open-sourcing soon — vista8 · 2026-11-09
- Multi-agent AI sales pipeline: 6 agents from lead discovery to supervisor review — atullchaurasia · 2026-11-09
- antirez: devs refusing AI coding help are visibly falling behind — antirez · 2026-11-09
- Coding agent pauses to ask its human: 'I need you to create one Hugging Face key' — Sauers_ · 2026-11-09
- Real-world coding agent benchmark: best model fully solves only 26.4% of tasks — emmanuelvivier · 2026-11-09