What Engineers Still Need to Know When AI Agents Code

Why it matters

A developer recounts how continuous integration flagged an AI-generated file for cyclomatic complexity after Codex wrote entire components on single lines. After ruling out a suspicious npm dependency, the author formatted the code, inspected the logic, and concluded the complexity was justified by business requirements, even if slightly over-engineered. Codex then refactored the file into multiple files, satisfying SonarQube so the build passed. The author frames this small incident as a preview of a broader shift in software ownership: as agents produce code faster than humans can absorb, the traditional path to system ownership through years of hands-on writing and reading code erodes. The piece contrasts optimists, who let autonomous agents build while they supervise from a distance, with skeptics, who warn of agents producing unmanageable, grotesque systems that later agents discard. The author argues that owning a system will increasingly mean knowing just enough to detect when something is wrong and being willing to investigate deeply when necessary, rather than having written the code by hand.