Andrej Karpathy’s term “agentic engineering” distinguishes disciplined, spec-driven, AI-agent-assisted development from “vibe coding,” which means accepting AI output without review.
Key Takeaways
Vibe coding is defined by not reviewing diffs; useful only for MVPs, prototypes, and one-off scripts where correctness is irrelevant.
Agentic engineering requires a design doc before prompting, per-diff code review, and a test suite the agent loops against until passing.
Tests are the critical differentiator: without them, agents confidently ship broken code; with them, agents become reliable systems.
Agentic engineering benefits senior engineers disproportionately because reviewing and correcting AI output requires strong fundamentals.
The author argues better specs, more comprehensive tests, and cleaner architecture all amplify AI output quality, raising the bar for engineering craft.
