If you are good at code review, you will be good at using AI agents (www.seangoedecke.com)

The author argues that effectively using AI coding agents (Claude Code, Codex, Copilot) is fundamentally a code‑review skill: LLMs are excellent at churning out code but lack engineering judgment, so left unchecked they often take architectural wrong turns, overengineer, or reverse‑engineer fragile solutions. Concrete examples include a VicFlora Offline PWA where Codex tried to repro the frontend instead of simply pulling the raw dichotomous‑key data, and a learning‑app prototype where agents insisted on building a full background‑job system (job entities, polling) when a simple non‑blocking frontend request would suffice. These misdirections cost time, tokens, and long‑term maintainability. For practitioners and teams this means the dominant working model remains “centaur chess”: a skilled human guiding an agent. The most valuable review is structural—bringing broader codebase context, preferring reuse over new subsystems, and catching architectural dead ends early—rather than line‑level nitpicking or rubber‑stamping. Engineers who can spot when an agent’s approach is the wrong place for the code will extract far more value from AI tooling; those who can’t will accrue wasted effort and complexity. The piece closes by noting agent capabilities have improved like a junior gaining experience, but they still require close human supervision rather than blind delegation.