Vibe coding and agentic engineering are getting closer than I'd like
Simon Willison now trusts AI agents to ship production code without reviewing every line.
Simon Willison, a prominent voice in AI-assisted programming, recently shared a disturbing realization on the Heavybit podcast: the line between 'vibe coding' and 'agentic engineering' is blurring in his own work. Vibe coding, coined by Andrej Karpathy, involves using AI to generate code without understanding or reviewing it—ideal for personal tools where bugs only hurt the user. Agentic engineering, which Willison advocates, relies on a professional engineer’s deep experience to guide AI tools, reviewing every line for security, maintainability, and performance.
But as agents like Claude Code prove extremely reliable for routine tasks—like building a JSON API endpoint with SQL queries—Willison finds himself skipping code review even for production systems. He draws an analogy: at large organizations, teams trust other teams’ black-box services without reading their code. Now he treats AI agents similarly, but feels guilty because humans are accountable. This convergence raises urgent questions about responsibility and code quality as AI coding tools become indistinguishable from human collaborators.
- Willison distinguishes vibe coding (blind trust in AI, no code review) from agentic engineering (using AI with professional oversight).
- He now skips reviewing routine agent-generated code for production, comparing it to trusting another engineering team’s black-box service.
- The convergence worries him because human accountability can’t be delegated to AI, even as agents grow more reliable.
Why It Matters
As AI coding agents near human-level reliability, even cautious experts must rethink when trust overrides code review.