Simon Willison Admits Vibe Coding and Agentic Engineering Have Merged
The post hit 601 points on HN, the top story of the day. Simon Willison has spent two years drawing a hard line: vibe coding is throwaway non-expert generation that you do not review, agentic engineering is professional AI-assisted development with full review and quality gates. The line just collapsed in his own practice and he wrote a confession about it.
The trigger: he has stopped reviewing Claude Code's output for production work. "If you ask Claude Code to build a JSON API endpoint, it is just going to do it right." So he just runs the code. Same behavior he criticized for two years.
His resolution is the interesting part. He treats Claude Code as a black-box service β same trust model as depending on AWS or Stripe. But he flags the asymmetry: AWS has accountability, Claude Code does not have a professional reputation that can be ruined. So you get vibe coding's reduced friction with agentic engineering's stakes, and the quality bar is just "does this work in production."
The structural read: when models cross a reliability threshold, the review economics flip. Reviewing a 200-line LLM patch line by line costs more than just running it and rolling back if it breaks. The Anthropic Skills movement, addyosmani/agent-skills (30k+ stars), and harness-safety cluster (Mendral / Rosentic / Mindra / Intuned / Tilde / AgentTrust) are all reactions to the same shift. Simon's post is the moment that shift gets named in public by someone who held the line longest.
Source: https://simonwillison.net/2026/May/6/vibe-coding-and-agentic-engineering/
← Back to all articles
The trigger: he has stopped reviewing Claude Code's output for production work. "If you ask Claude Code to build a JSON API endpoint, it is just going to do it right." So he just runs the code. Same behavior he criticized for two years.
His resolution is the interesting part. He treats Claude Code as a black-box service β same trust model as depending on AWS or Stripe. But he flags the asymmetry: AWS has accountability, Claude Code does not have a professional reputation that can be ruined. So you get vibe coding's reduced friction with agentic engineering's stakes, and the quality bar is just "does this work in production."
The structural read: when models cross a reliability threshold, the review economics flip. Reviewing a 200-line LLM patch line by line costs more than just running it and rolling back if it breaks. The Anthropic Skills movement, addyosmani/agent-skills (30k+ stars), and harness-safety cluster (Mendral / Rosentic / Mindra / Intuned / Tilde / AgentTrust) are all reactions to the same shift. Simon's post is the moment that shift gets named in public by someone who held the line longest.
Source: https://simonwillison.net/2026/May/6/vibe-coding-and-agentic-engineering/
Comments