How do you stop an AI agent from overspending on GPUs?

A prompt is not a limit. Cap every run, put a blocking budget around agent work, scope its credentials and approve writes. Setup for Claude Code and Codex.