Vibe Managing (ag404labs.com)

After six months running AI-generated code in production across banking, infra, and audited systems, the author reports a clear shift: AI has made typing cheap and scalable, but it hasn’t removed the need for human judgment. Tools can scaffold CRUD, draft Terraform, and produce thousands of plausible lines quickly—often 80–95% correct—but that speed creates volume: human review load jumping from ~5 PRs/week to ~50, each looking right at a glance. One missed Terraform constraint that passed validation opened a regulated system to the internet (alerts caught it), underscoring that responsibility still rests with the human approver. For the AI/ML community, the practical implication is that developers are being involuntarily promoted into managerial roles: triage, risk ownership, and decision-making matter more than syntax. The rules of safe deployment remain the same—don’t let juniors or AI merge to prod, review logic not commas, automate linting/tests/CI/security scans, and pair AI with humans. Training should shift from teaching syntax to teaching specification, defining “done,” and owning blast radius. In short: AI supplies infinite junior-like output; humans must scale judgment and governance to prevent small mistakes from becoming major breaches.