What if Code Review Happened Before the Code was Written?

AI agents can generate code faster than teams can review it, creating a new bottleneck in the SDLC. A proposed solution is spec-driven verification: humans review a detailed spec before any code is written, an agent implements it, and a second agent verifies the output against pre-agreed acceptance criteria. An experiment using Claude Code implemented a full-stack feature (hierarchical per-repo config system) in 5 hours, producing ~6k lines of code with 40% tests. A verification agent checked 65 acceptance criteria in 6 minutes (60 passed, 4 failed, 1 partial). Key findings: spec reviews catch design-level issues early, code reviews still catch convention violations, and agents tend to create patterns that contradict existing codebase conventions without explicit guidance.