Willison: The line between vibe coding and agentic engineering is blurring

Simon Willison reflects on a shift in his own AI-assisted coding practice. He previously drew a sharp line between ‘vibe coding’ — accepting AI output without reading it, fine for personal tools but reckless for production — and ‘agentic engineering,’ where an experienced developer leans on AI while still applying professional judgment about security, maintainability, and operations. As coding agents like Claude Code have grown more reliable on routine tasks, he admits he no longer reviews every line they produce, even for production systems, and compares the relationship to consuming an internal service from another team: trust the interface, dig into the implementation only when something breaks. He flags the normalization-of-deviance risk this creates.

The productivity jump exposes second-order problems. Evaluating open-source projects by commit count, README polish, and test coverage no longer works because an agent can fabricate all of those signals in half an hour; actual sustained use by the author has become a stronger quality signal. Bottlenecks across the software lifecycle, both upstream design processes and downstream review, were calibrated for a world where a developer wrote a few hundred lines a day, and that assumption is breaking.

Willison remains unworried about his career, framing the tools as amplifiers of existing expertise rather than replacements, and endorses Matthew Yglesias’s view that most people would rather buy software from professionals using AI than vibe-code their own.