Vibe engineering (simonwillison.net)

Author proposes “vibe engineering” as a name for the disciplined, accountable end of the AI-assisted development spectrum—contrasting with “vibe coding,” the fast, loose, prompt-driven approach. The rise of coding agents (Claude Code — Feb 2025, OpenAI Codex CLI — April, Gemini CLI — June) that iteratively write, test and modify code has made LLMs genuinely useful for non-toy, production problems. Experienced engineers are now running parallel agents to tackle multiple tasks, but getting reliable results requires active human oversight, planning and specialised workflows rather than blind prompt-and-forget usage. The piece argues that LLMs amplify good engineering practices: robust automated test suites (test-first loops), advance planning and specs, comprehensive documentation to work around context-window limits, strict version control and git fluency (agents handle git history and bisect well), CI/CD, preview environments, disciplined code review, manual QA, and strong research/estimation skills. Practically, “vibe engineering” means designing agentic loops, defining success criteria, and deciding what to outsource vs. what to keep manual—roles that map to senior engineering responsibilities. The net effect: AI tools accelerate work but raise the bar for engineering rigor, making expertise more valuable rather than obsolete.