The Boundary of AI Coding Is Blurring

Simon Willison said on Heavybit's High Leverage podcast that vibe coding and agentic engineering are increasingly overlapping in his own work. He used to distinguish the former as a method of receiving results with barely looking at the code, and the latter as a professional's use of AI that takes responsibility for security, maintenance, and operations, but he now sees that the boundary is no longer as clear as it used to be. He also said that while it might be fine for personal tools, it's irresponsible to build production software that handles other people's information through vibe coding.

• As Claude Code reliably handles simple JSON API and SQL work, he no longer reviews some production code line by line like he used to.
• Just as he doesn't read all of another team's image resize service in-house, he's come to treat agents as semi-black boxes that he only looks into when necessary.
• This has also raised concerns about normalization of deviance — the more often unreviewed code turns out to be correct, the more it can lead to overconfidence at the wrong moment later on.

Evaluation criteria have also changed. In the past, a repository with many commits and a well-maintained README and tests offered some way to gauge quality, but now anyone can produce a similar-looking appearance in a short time. So he said he now trusts a tool that has actually been used for a few weeks more than its outward appearance.

The rise in production speed has also shifted the bottleneck. When a flow that used to produce 200 lines a day shifts to 2,000 lines, the entire design and development process has to be redesigned, and the cost of correcting a wrong decision also drops. He believes AI tools only make a 25-year veteran engineer faster, without making software itself any less difficult, and that's why he doesn't think careers are over. He summed it up by saying there's no point in producing lower quality faster, and that the goal is to produce higher quality faster. He added that enterprise software also needs solutions with an actual track record of use — like a proven CRM that at least two large companies have used for more than 6 months.