Developer Shares Workflow to Prevent Runaway AI Coding Agent Loops

A software developer published a reflective post-mortem after an AI coding agent ran unchecked in a loop for hours on a Friday afternoon, burning through tokens with duplicate tool calls. The core problem was treating free remote compute as unlimited, while the agent lacked a defined stop rule and ran with unclean laptop state including sensitive files like .env. The author proposes a structured workflow requiring a clean git worktree, an isolated remote workspace with only allowlisted files, and strict budget limits on steps and runtime. A session receipt must be opened before the first prompt and closed with a hashed diff once the agent stops, preserving a replayable audit trail. The accompanying Node.js script enforces three control gates — budget, isolation, and receipt — to prevent similar failures.