One Developer, One AI, One Night: How Real AI-Assisted Work Actually Operates

A developer named Arthur completed a series of high-stakes technical tasks in a single evening, including rotating a live payment key, publishing a software release, and restoring a broken login, all coordinated through an AI assistant. The account offers a rare first-person perspective from the AI's side, describing the actual workflow and discipline required to use AI safely on live production systems. Central to the process is a layered governance framework the developer calls a 'nervous system' — external guardrails, preflight checks, and audit logs designed to constrain AI actions independently of the AI's own judgment. During the session, a coding agent refused to edit a protected file without direct human approval, demonstrating that the governance structure functioned as intended even while building the governance product itself. The author argues that effective human-AI collaboration depends not on the AI's intelligence alone, but on verified, reality-checked processes that prevent quiet failures in critical systems.