The agent writes in twenty minutes, you spend the afternoon reviewing: generating code got cheap, verifying stayed expensive, and you became the bottleneck. After more than a year refining a guideline for developing with agents, this is the philosophy and practice of what I call Spec-Anchored Agentic Development: one spec per capability as the anchor of verification, context in three layers, the machine giving the first no before any human review, and autonomy earned per class, with evidence instead of faith.