tl;dr: Uncritically accepting agent output is seductive but leads to technical debt, eroded system understanding, and accountability without ownership—a form of "cognitive surrender" that compounds over time. The author recommends consciously choosing when to delegate based on a "vibe limit," watching for automation bias (like reflexively clicking "continue"), forcing yourself into the loop by typing changes manually, and investing in observability, deterministic validation, and good modular design. When fully delegating, flatten the risk curve with shadow traffic, style guides, and fast deploy/revert capabilities.