# The Claude Code Idempotency Test That Prevents Runaway Agent Costs > Source: > Published: 2026-08-05 06:29:35+00:00 Member-only story # The Claude Code Idempotency Test That Prevents Runaway Agent Costs A single scheduling flaw can double your API bill — here is how to choose between loops, skills, subagents, and workflows without burning budget. A hidden bug in Claude Code’s scheduling layer is quietly re-firing single-run commands every time a background task wakes up. If that command performs a database write, opens a pull request, or triggers an expensive agentic search, it won’t just duplicate your work — it will bill you twice. Most guides treat `/loop` , `/goal` , and `/schedule` as separate tricks, while filing skills versus subagents away as a completely unrelated architectural choice. They miss the core connection: a single design gap in `ScheduleWakeup` links all of them together. If you are using free version of Medium, [Please consider]use link to access this article for free.[, share this article with your friends to support my work. Thank you.]following and subscribing Before you automate any prompt, you must answer one non-negotiable question: **is this prompt idempotent to re-fire?** In other words, if the scheduler re-executes your prompt halfway through the run, will the final state remain safe, clean, and financially sane? **In this article you’ll learn**