AI Budgets Require Code Enforcement, Not Just Prompt Instructions

The article argues that spending limits written into AI prompts are ineffective because they are mere text that the model can ignore or override. A technical failure is cited where a self-hosted agent repeatedly paid for a failing purchase due to the lack of an external authorization check. The author concludes that real budget enforcement must exist outside the model as a scoped and digitally signed authorization verified by a server. This prevents issues like repetitive payment loops, prompt injection, and child agents exceeding their budgets.