Agent-driven “AI” promises effortless coding, but the economics of context make that promise fragile. In practice, Claude Code often loses the thread, breaks working systems, and swaps upbeat rhetoric for results—especially beyond toy prototypes. Delegating to sub-agents helps little when the lead forgets the plan. This essay maps the failure modes and argues that, until context and rigor improve, uncritical adoption will erode quality even as delivery speeds up.